WO2024031701A1 - 数据获取方法与装置 - Google Patents

数据获取方法与装置 Download PDF

Info

Publication number
WO2024031701A1
WO2024031701A1 PCT/CN2022/112299 CN2022112299W WO2024031701A1 WO 2024031701 A1 WO2024031701 A1 WO 2024031701A1 CN 2022112299 W CN2022112299 W CN 2022112299W WO 2024031701 A1 WO2024031701 A1 WO 2024031701A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
information
model
csi
unit
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2022/112299
Other languages
English (en)
French (fr)
Inventor
王昕�
张群
孙刚
张健
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fujitsu Ltd
Original Assignee
Fujitsu Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Fujitsu Ltd filed Critical Fujitsu Ltd
Priority to CN202280098380.XA priority Critical patent/CN119586082A/zh
Priority to PCT/CN2022/112299 priority patent/WO2024031701A1/zh
Priority to EP22954653.6A priority patent/EP4572243A4/en
Publication of WO2024031701A1 publication Critical patent/WO2024031701A1/zh
Priority to US19/029,817 priority patent/US20250192849A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station
    • H04B7/0613—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission
    • H04B7/0615—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal
    • H04B7/0619—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal using feedback from receiving side
    • H04B7/0621—Feedback content
    • H04B7/0626—Channel coefficients, e.g. channel state information [CSI]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/0413—MIMO systems
    • H04B7/0456—Selection of precoding matrices or codebooks, e.g. using matrices antenna weighting
    • H04B7/0478—Special codebook structures directed to feedback optimisation
    • H04B7/048—Special codebook structures directed to feedback optimisation using three or more PMIs
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/0413—MIMO systems
    • H04B7/0456—Selection of precoding matrices or codebooks, e.g. using matrices antenna weighting
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0254—Channel estimation channel estimation algorithms using neural network algorithms
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station
    • H04B7/0613—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission
    • H04B7/0615—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal
    • H04B7/0619—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal using feedback from receiving side
    • H04B7/0658—Feedback reduction

Definitions

  • the embodiments of this application relate to the field of communication technology.
  • Multi-antenna technology is widely used in LTE, LTE-A and 5G NR systems.
  • massive antenna technology massive MIMO
  • massive MIMO massive MIMO
  • 5G-Advanced stage and the 6G stage massive MIMO technology will be more widely used.
  • Large-scale and ultra-large-scale antenna technologies with enhanced performance are also the focus of research on next-generation mobile communication systems.
  • AI Artificial Intelligence
  • ML Machine Learning
  • the CSI is measured on the terminal device side, and the AI/ML model is used to generate CSI feedback information. After sending it to the network side through the air interface, the network side receives the CSI feedback information, and restore the original CSI through the corresponding AI/ML model.
  • the AI/ML model can reduce the CSI feedback overhead or improve the quality of feedback, thereby improving communication quality.
  • the information from the sending end (network device or terminal device) is processed by the AI/ML model, it is sent to the receiving end (terminal device or network device) through the air interface.
  • the receiving end uses the model corresponding to the AI/ML model of the sending end.
  • this type of AI/ML model can usually be called a two-sided model.
  • the pairwise attribute of the bilateral model requires the cooperation of network equipment and terminal equipment on both sides of the air interface in the training process, and network model training is performed through the air interface, which will cause huge air interface overhead.
  • embodiments of the present application provide a data acquisition method, wherein the AI/ML model includes an information generation unit located in the first device and an information reconstruction unit located in the second device.
  • the method include:
  • the second device acquires the first data X input to the information generation unit
  • the second device acquires second data Y corresponding to the first data X and output from the information generation section.
  • a data acquisition device configured in a second device; wherein the AI/ML model includes an information generation unit located in the first device and an information reconstruction unit located in the second device. part, the data acquisition device includes:
  • a first acquisition unit that acquires the first data X input to the information generation unit
  • a second acquisition unit acquires the second data Y corresponding to the first data X and output from the information generation unit.
  • a data acquisition method wherein the AI/ML model includes an information generation part located in the first device and an information reconstruction part located in the second device, and the method includes:
  • the first device acquires the first data X output by the information reconstruction unit
  • the first device acquires the second data Y input to the information reconstruction section and corresponding to the first data X.
  • a data acquisition device configured in a first device; wherein the AI/ML model includes an information generation unit located in the first device and an information reconstruction unit located in the second device.
  • the device includes:
  • a third acquisition unit that acquires the first data X output by the information reconstruction unit
  • a fourth acquisition unit acquires the second data Y input to the information reconstruction unit and corresponding to the first data X.
  • a communication system in which the AI/ML model includes a CSI generation unit located in the terminal device and a CSI reconstruction unit located in the network device, and the system includes:
  • a network device that acquires a specific codebook vector input to the CSI generation unit; and acquires CSI bit information output from the CSI generation unit corresponding to the codebook vector;
  • a terminal device acquires a codebook vector output by the CSI reconstruction unit; and acquires CSI bit information input to the CSI reconstruction unit and corresponding to the codebook vector.
  • the second device acquires the first data DataY.
  • suitable data can be obtained at a relatively low cost, thereby supporting the bilateral network model architecture.
  • Figure 1 is a schematic diagram of a communication system according to an embodiment of the present application.
  • Figure 2 is an example diagram of the bilateral model according to the embodiment of the present application.
  • Figure 3 is another example diagram of the bilateral model according to the embodiment of the present application.
  • Figure 4 is another example diagram of the bilateral model according to the embodiment of the present application.
  • Figure 5 is another example diagram of the bilateral model according to the embodiment of the present application.
  • Figure 6 is a schematic diagram of the data acquisition method according to the embodiment of the present application.
  • Figure 7 is a schematic diagram of data in the embodiment of the present application.
  • Figure 8 is a schematic diagram of joint training according to the embodiment of the present application.
  • Figure 9 is a schematic diagram of separation training according to an embodiment of the present application.
  • Figure 10 is a schematic diagram after separation training according to the embodiment of the present application.
  • Figure 11 is a schematic diagram of the application of a bilateral model in gNB and UE according to the embodiment of the present application.
  • Figure 12 is a schematic diagram of the data acquisition method according to the embodiment of the present application.
  • Figure 13 is an example diagram of constructing a data set according to the embodiment of the present application.
  • Figure 14 is a schematic diagram of the data acquisition method according to the embodiment of the present application.
  • Figure 15 is an example diagram of data obtained by the CSI reconstruction unit according to the embodiment of the present application.
  • Figure 16 is an example diagram of training performed by the CSI reconstruction unit according to the embodiment of the present application.
  • Figure 17 is another example diagram of data obtained by the CSI reconstruction unit according to the embodiment of the present application.
  • Figure 18 is another example diagram of training performed by the CSI reconstruction unit according to the embodiment of the present application.
  • Figure 19 is another example diagram of training performed by the CSI reconstruction unit according to the embodiment of the present application.
  • Figure 20 is a signaling diagram of the CSI reconstruction unit training according to the embodiment of the present application.
  • Figure 21 is a schematic diagram of the data acquisition method according to the embodiment of the present application.
  • Figure 22 is an example diagram of data obtained by the CSI generation unit according to the embodiment of the present application.
  • Figure 23 is another example diagram of training performed by the CSI generation unit according to the embodiment of the present application.
  • Figure 24 is a signaling schematic diagram of training by the CSI generation unit according to the embodiment of the present application.
  • Figure 25 is a schematic diagram of the data acquisition device according to the embodiment of the present application.
  • Figure 26 is another schematic diagram of the data acquisition device according to the embodiment of the present application.
  • Figure 27 is a schematic diagram of network equipment according to an embodiment of the present application.
  • Figure 28 is a schematic diagram of a terminal device according to an embodiment of the present application.
  • the terms “first”, “second”, etc. are used to distinguish different elements from the title, but do not indicate the spatial arrangement or temporal order of these elements, and these elements should not be used by these terms. restricted.
  • the term “and/or” includes any and all combinations of one or more of the associated listed terms.
  • the terms “comprises,” “includes,” “having” and the like refer to the presence of stated features, elements, elements or components but do not exclude the presence or addition of one or more other features, elements, elements or components.
  • the term “communication network” or “wireless communication network” may refer to a network that complies with any of the following communication standards, such as Long Term Evolution (LTE, Long Term Evolution), Long Term Evolution Enhanced (LTE-A, LTE- Advanced), Wideband Code Division Multiple Access (WCDMA, Wideband Code Division Multiple Access), High-Speed Packet Access (HSPA, High-Speed Packet Access), etc.
  • LTE Long Term Evolution
  • LTE-A Long Term Evolution Enhanced
  • LTE-A Long Term Evolution Enhanced
  • WCDMA Wideband Code Division Multiple Access
  • High-Speed Packet Access High-Speed Packet Access
  • communication between devices in the communication system can be carried out according to any stage of communication protocols, which may include but are not limited to the following communication protocols: 1G (generation), 2G, 2.5G, 2.75G, 3G, 4G, 4.5G and 5G. , New Wireless (NR, New Radio), future 6G, etc., and/or other communication protocols currently known or to be developed in the future.
  • Network device refers to a device in a communication system that connects a terminal device to a communication network and provides services to the terminal device.
  • Network equipment may include but is not limited to the following equipment: base station (BS, Base Station), access point (AP, Access Point), transmission and reception point (TRP, Transmission Reception Point), broadcast transmitter, mobile management entity (MME, Mobile Management Entity), gateway, server, wireless network controller (RNC, Radio Network Controller), base station controller (BSC, Base Station Controller), etc.
  • the base station may include but is not limited to: Node B (NodeB or NB), evolved Node B (eNodeB or eNB), 5G base station (gNB), IAB host, etc.
  • NodeB Node B
  • eNodeB or eNB evolved Node B
  • gNB 5G base station
  • IAB host etc.
  • it may also include Remote Radio Head (RRH, Remote Radio). Head), remote wireless unit (RRU, Remote Radio Unit), relay or low-power node (such as femeto, pico, etc.).
  • RRH Remote Radio Head
  • RRU Remote Radio Unit
  • relay or low-power node such as femeto, pico, etc.
  • base station may include some or all of their functions, each of which may provide communications coverage to a specific geographic area.
  • the term "cell” may refer to a base station and/or its coverage area, depending on the context in which the term is used.
  • the term "user equipment” (UE, User Equipment) or “terminal equipment” (TE, Terminal Equipment or Terminal Device) refers to a device that accesses a communication network through a network device and receives network services.
  • Terminal equipment can be fixed or mobile, and can also be called mobile station (MS, Mobile Station), terminal, subscriber station (SS, Subscriber Station), access terminal (AT, Access Terminal), station, etc.
  • the terminal equipment may include but is not limited to the following equipment: cellular phone (Cellular Phone), personal digital assistant (PDA, Personal Digital Assistant), wireless modem, wireless communication equipment, handheld device, machine-type communication equipment, laptop computer, Cordless phones, smartphones, smart watches, digital cameras, and more.
  • cellular phone Cellular Phone
  • PDA Personal Digital Assistant
  • wireless modem wireless communication equipment
  • handheld device machine-type communication equipment
  • laptop computer Cordless phones
  • Cordless phones smartphones, smart watches, digital cameras, and more.
  • the terminal device can also be a machine or device for monitoring or measuring.
  • the terminal device can include but is not limited to: Machine Type Communication (MTC) terminals, Vehicle communication terminals, device-to-device (D2D, Device to Device) terminals, machine-to-machine (M2M, Machine to Machine) terminals, etc.
  • MTC Machine Type Communication
  • D2D Device to Device
  • M2M Machine to Machine
  • network side refers to one side of the network, which may be a certain base station or may include one or more network devices as above.
  • user side or “terminal side” or “terminal device side” refers to the side of the user or terminal, which may be a certain UE or may include one or more terminal devices as above.
  • device can refer to network equipment or terminal equipment.
  • Figure 1 is a schematic diagram of a communication system according to an embodiment of the present application, schematically illustrating a terminal device and a network device as an example.
  • the communication system 100 may include a network device 101 and terminal devices 102 and 103.
  • Figure 1 only takes two terminal devices and one network device as an example for illustration, but the embodiment of the present application is not limited thereto.
  • eMBB enhanced mobile broadband
  • mMTC massive machine type communication
  • URLLC Ultra-Reliable and Low -Latency Communication
  • Figure 1 shows that both terminal devices 102 and 103 are within the coverage of the network device 101, but the application is not limited thereto. Neither of the two terminal devices 102 and 103 may be within the coverage range of the network device 101, or one terminal device 102 may be within the coverage range of the network device 101 and the other terminal device 103 may be outside the coverage range of the network device 101.
  • the high-level signaling may be, for example, Radio Resource Control (RRC) signaling; for example, it is called an RRC message (RRC message), and for example, it includes MIB, system information (system information), and dedicated RRC message; or it is called RRC IE (RRC information element).
  • RRC Radio Resource Control
  • high-level signaling may also be MAC (Medium Access Control) signaling; or it may be called MAC CE (MAC control element).
  • RRC Radio Resource Control
  • RRC message RRC message
  • MIB system information (system information), and dedicated RRC message
  • RRC IE RRC information element
  • high-level signaling may also be MAC (Medium Access Control) signaling; or it may be called MAC CE (MAC control element).
  • MAC CE Medium Access Control
  • one or more AI/ML models can be configured and run in network devices and/or terminal devices.
  • the AI/ML model can be used for various signal processing functions of wireless communication, such as CSI estimation and reporting, beam management, beam prediction, etc.; this application is not limited thereto.
  • FIG. 2 is an example diagram of the bilateral model in the embodiment of the present application.
  • part of the AI/ML model is at the sending end TX, and the other part is at the receiving end RX.
  • source information X is input into the model to generate information Y, which is transmitted to the receiving end.
  • information Y’ transmitted through the channel is input into the model to generate recovery information X’.
  • the above information satisfies the following formula 1:
  • f( ⁇ ) is the transfer function of the TX side neural network
  • X is the source information of the neural network at TX
  • Y is its output signal
  • h( ⁇ ) is the channel response coefficient between TX and RX
  • n is Gaussian noise
  • g( ⁇ ) is the transfer function of the RX side neural network
  • Y' is the input signal of the neural network at RX
  • X' is its output signal.
  • the loss function makes X' as close to X as possible, that is, ⁇ X-X' ⁇ 0.
  • the loss function can use cosine similarity, NMSE or MSE as the metric.
  • the TX neural network and RX neural network are deployed on the TX side (such as the UE side) and the RX side (such as the gNB side) respectively, and communicate with each other through the air interface.
  • the bilateral model can be applied to CSI generation and reporting.
  • FIG 3 is another example diagram of the bilateral model according to the embodiment of the present application.
  • the bilateral model consists of a CSI generation part (also called CSI encoder) on the terminal device side and a CSI reconstruction part (also called CSI decoder) on the network device side.
  • CSI generation part also called CSI encoder
  • CSI reconstruction part also called CSI decoder
  • the input of the CSI generation unit is channel coefficient information, which can be a channel coefficient vector or matrix generated by channel estimation, or a channel feature vector or singular vector generated by matrix decomposition of the above channel coefficient matrix.
  • the output of the CSI generation unit is a CSI bit sequence formed through feature extraction, information compression, and quantization. This bit sequence is carried on the uplink control channel or the uplink data channel as part of the uplink control information (UCI), and is sent to the network side.
  • UCI uplink control information
  • the network device extracts the corresponding CSI bit sequence from the received UCI and inputs it to the CSI reconstruction unit.
  • the output of the CSI reconstruction unit is a reconstructed channel coefficient vector or matrix, or a channel feature vector or a singular vector, whichever corresponds to the input of the CSI generation unit.
  • the bilateral model can also replace traditional signal processing methods and be used in various pairwise processing modules in communication systems.
  • FIG. 4 is another example diagram of the bilateral model according to the embodiment of the present application.
  • source coding and decoding source coding/decoding
  • channel coding and decoding channel coding/decoding
  • channel modulation and demodulation modulation/demodulation
  • multi-antenna module mimo/ de-mimo specifically, it can include paired signal processing at both the sending and receiving ends for mimo multi-data stream multiplexing, mimo space-time coding, mimo precoding/beamforming, etc., and other sending ends from multiple data streams to multiple antenna ports.
  • the signal processing processes or functions of the transmitter and receiver at both ends of the communication can be merged, and a neural network model can be introduced to implement the original signal processing.
  • FIG. 5 is another example diagram of the bilateral model according to the embodiment of the present application.
  • the sending and receiving ends can be equipped with an information generation unit at the sending end and an information reconstruction unit at the receiving end, so that the bilateral model can be used to perform paired signal processing.
  • the construction and upgrade of this bilateral model need to be completed through joint training, that is, the information generation department and the information reconstruction department need to work together to complete joint initial training, joint fine tuning, joint retraining, and joint model upgrade.
  • All joint training related processes of the above-mentioned bilateral models include offline operations or online operations.
  • one side can transmit its matching model to the other side.
  • the network side transmits the CSI generation unit that matches the CSI reconstruction unit to the terminal side through the air interface, and the terminal side uses the CSI generation unit to work after receiving it.
  • this requires the terminal side to have the ability to receive and upgrade AI/ML, and the requirements for its hardware and software are very high.
  • the solution for transmitting and using AI/ML models in network equipment and terminal equipment has intellectual property issues beyond device capabilities and collaboration issues between other manufacturers. Therefore, the solution to achieve bilateral model pairing through model transfer will face many problems.
  • the AI/ML model on the terminal side or the network side can be updated so that the models on both sides can be paired and work jointly.
  • the AI/ML models on one or both sides need to be jointly trained.
  • Model training requires a data set, and for supervised learning, a labeled data set is also required.
  • joint training brings very large air interface overhead, including the overhead of outputting related data sets, the overhead of input and output corresponding to the back-and-forth iteration process, etc., which is difficult to implement over the air interface of mobile communications. Support this process.
  • the embodiments of this application propose a solution to use shared paired data sets to achieve independent training, and provide methods and devices for reducing the size of the shared data set and reducing the air interface overhead of acquiring data.
  • Joint training includes the process in which the two sides of the bilateral model need to be jointly trained (including initial training, retraining, online training, offline training, fine tuning, etc., and also includes the necessary steps in training the model) sub-process of model verification and model testing).
  • Separate training is also called separate training or independent training, including the process in which the two sides of the bilateral model can be trained independently (including initial training, retraining, online training, offline training, fine tuning, etc., and also includes the training model Necessary model verification and model testing sub-processes).
  • the embodiment of the present application provides a data acquisition method, which is explained from the second device side.
  • the AI/ML model includes an information generation part located in the first device and an information reconstruction part located in the second device.
  • the second device may be a network device and the first device is a terminal device; the second device may also be a terminal device and the first device is a network device.
  • Figure 6 is a schematic diagram of the data acquisition method according to the embodiment of the present application. As shown in Figure 6, the method includes:
  • the second device obtains the first data X input to the information generation unit;
  • the second device obtains the second data Y corresponding to the first data X and output from the information generation unit.
  • AI/ML models can be run separately for different signal processing functions.
  • the AI/ML model reported by CSI can have different model group identifiers, model identifiers, and version identifiers.
  • AI/ML models for beam management may have additional model group identifiers, model identifiers, and version identifiers.
  • the AI/ML model is a bilateral model with a model identification and a version identification; the information generation part and the information reconstruction part of the same AI/ML model use the same model identification and version identification, and the information generation part
  • the information reconstruction part and the information reconstruction part have different sub-identities.
  • a bilateral model has a model identifier (which can be a one-level identifier or a multi-level identifier) and a version identifier, which can distinguish two different bilateral models (such as the number of layers, number of nodes, hyperparameters, input and output signal formats, preset processing and post-processing configuration, etc.).
  • the information generation department and information reconstruction department of the same bilateral model have the same model identification and version identification.
  • a sub-identification can be used to distinguish whether it is an information generation unit or an information reconstruction unit.
  • the model identification is the first-level identification XXXXX (single-level model identification).Vxx (version identification);
  • the model identification is a multi-level identification: first level XXX-second level XXXX-nth level XXXX (multi-level model identification).Vxxx (version identification).
  • a 1-bit sub-identifier can be used to distinguish the generation part and the reconstruction part.
  • the information generation part is: model identification.version identification.0; the information reconstruction part is: model identification.version identification.1.
  • Figure 7 is a schematic diagram of the data in the embodiment of the present application. As shown in Figure 7, for a bilateral model, although it is deployed separately on the sending side and the receiving side, the input and output of the information generation part correspond to the output of the information reconstruction part. and input. ideal,
  • Table 1 shows examples of data for embodiments of the present application.
  • the information generation part and the information reconstruction part of a bilateral model corresponding to a certain model identification have the same (or approximately) dual training data set. That is, the input data set ⁇ X ⁇ and the corresponding output data set ⁇ Y ⁇ of the information generation part can be shared with the information reconstruction part, and the information reconstruction part inputs ⁇ Y ⁇ and outputs ⁇ X ⁇ . Therefore, this feature can be used for separation training,
  • the second device inputs the second data Y into the information reconstruction part, and uses the first data X as label data to train the information reconstruction part.
  • the information generation part in the first device and the information reconstruction part in the second device use the first data X and the second data Y to perform training respectively;
  • the output is the second data Y or data that is similar to the second data Y;
  • the information generation unit that has been trained in the second device When the second data Y is input, the information reconstruction unit outputs the first data X or data approximate to the first data X.
  • the information reconstruction part in the network device and the information generation part of the terminal device use the same paired data sets (common paired datasets) ⁇ X ⁇ and ⁇ Y ⁇ for training respectively. .
  • the output is ⁇ X ⁇ or approximately ⁇ X ⁇ ;
  • the output is ⁇ Y ⁇ or approximately for ⁇ Y ⁇ ;
  • the formats of X and Y are predefined or configured, such as symbol format, number of bits, etc.
  • the first data X and the second data Y are paired data sets; the paired data sets have model identification information. Among them, one first data X can correspond to multiple second data Y.
  • one X can correspond to multiple Ys (such as Y1, Y2), and the formats of Y1 and Y2 are predefined or configured.
  • the paired data set is: ⁇ X ⁇ Y1 ⁇ , ⁇ X ⁇ Y2 ⁇ , or ⁇ X ⁇ Y1 ⁇ Y2 ⁇ , for example, it can be on multiple first devices and one second device used in.
  • Figure 8 is a schematic diagram of joint training according to the embodiment of the present application.
  • the model of the information generation part needs to be jointly trained with the model of the information reconstruction part, after multiple iterations, backward propagation, parameter upgrades and other steps. , complete the training.
  • Figure 9 is a schematic diagram of separation training according to an embodiment of the present application.
  • the first data X and the second data Y are prepared in advance.
  • X and Y are in a paired relationship.
  • Multiple Xs and corresponding multiple Ys constitute a paired data set, where X
  • the format of Y is predefined or configured, such as symbol format, number of bits, etc.
  • One X can correspond to multiple Ys (such as Y1, Y2), and the formats of Y1 and Y2 are predefined or configured.
  • the paired data sets are: ⁇ X ⁇ Y1 ⁇ , ⁇ X ⁇ Y2 ⁇ , or ⁇ X ⁇ Y1 ⁇ Y2 ⁇ .
  • the prepared data set can be extracted from various typical channels through simulation analysis.
  • the paired data set is: ⁇ X ⁇ Y1 ⁇ , ⁇ X ⁇ Y2 ⁇ , or ⁇ X ⁇ Y1 ⁇ Y2 ⁇ , for example, it can be on multiple first devices and one second device used in.
  • ⁇ X ⁇ of the common paired data set is used as the input data of the UE side model part, and ⁇ Y ⁇ is used as the output reference data or label data.
  • the purpose of training is to make the input The output for every X is Y, or very close to Y.
  • the same data set is also shared with the gNB side model department.
  • ⁇ Y ⁇ is used as the input data of the gNB side model department
  • ⁇ X ⁇ is used as the output reference data or label data.
  • the purpose of training is to make the output when each Y is input after training. It's X, or very close to X.
  • the above paired data sets are used to train the UE side model part (for example, called the information generation part) and the gNB side model part (for example, called the information reconstruction part), based on the above training requirements, after the separate training is completed, When jointly working together, at least for the input data set ⁇ X ⁇ , the same data recovery effect as joint training can be achieved.
  • Figure 10 is a schematic diagram after separation training according to the embodiment of the present application. As shown in Figure 10.
  • the data set ⁇ X ⁇ sufficiently characterizes the characteristics of the wireless channel to which the model is applied, it will work very well when applied to the communication system.
  • the data set based on the channel update can be used to perform model retraining or fine tuning respectively to ensure that the model is correct.
  • Adaptability to new application environments. By continuously updating paired data sets and using this method to train paired bilateral models separately, the difficulty of joint training of bilateral models can be solved.
  • the above method can be used for initial training, retraining, offline training, online training, fine tuning and other training processes of the model.
  • an online and air interface-based approach can be used.
  • Figure 11 is a schematic diagram of the application of a bilateral model in gNB and UE according to the embodiment of the present application, in which many possible situations are considered.
  • model monitoring finds that performance is poor
  • the training process of the model can be started.
  • the model departments on both sides use the newly collected data to conduct interactive and independent training based on the above-mentioned principle of shared paired data sets.
  • the trained model can be used through the air interface. Use and performance monitoring, after the performance reaches the standard, it can be used as an updated model for model inference. This avoids problems faced by joint training and model transfer.
  • the model that has been trained and tested over the air interface can be registered with the channel model ID or new version ID, and can be uploaded to the core network, remote server, OTT server, or passed to other gNBs, etc.
  • the first data Corresponding data generated later.
  • the specific data set may be a data set that can be generated based on predefined rules corresponding to the current generation part model, or may be a data set common to all information generation part models, such as all possible channel information quantization vectors.
  • the second device obtains the first data X and/or the first data X from inside the second device or from outside the second device according to identification information related to the information generation part 2.
  • Data Y is a code that specifies the first data X and/or the first data X from inside the second device or from outside the second device according to identification information related to the information generation part 2.
  • the inside of the second device may include storage devices such as memory and disks; the outside of the second device may include devices connected to the second device, such as core networks, clouds, third-party servers, OTT servers, cloud storage devices, and so on.
  • storage devices such as memory and disks
  • the outside of the second device may include devices connected to the second device, such as core networks, clouds, third-party servers, OTT servers, cloud storage devices, and so on.
  • the identification information related to the information generation unit includes: model identification and/or version information corresponding to the information generation unit, and/or data configuration information of the model corresponding to the information generation unit.
  • the application is not limited to this, and may also be other identifiers, for example.
  • the second data Y is carried through a control channel or a data channel and transmitted over an air interface, or the second data Y is generated through a data index according to predefined rules, or the index of the second data Y is determined by The first device sends the information to the second device through the air interface.
  • the first data X is pre-stored in the second device, or the first data X is generated through a data index according to predefined rules, or the first data
  • the first device sends it to the second device through the air interface, or the index of the first data X is sent by the first device to the second device through the air interface.
  • the information generation unit or the information reconstruction unit performs initial training or offline training.
  • the information generation department and the information reconstruction department agree on the input data configuration and the corresponding output data configuration. Configuration can include the dimensions, format, etc. of the data. Then train separately, and the specific training methods are not limited.
  • the information generation part is trained first.
  • the model part paired with it can be used to train together during the process.
  • the information generation part shares all or part of its input data set and the corresponding final output data set to the reconstruction part.
  • the information reconstruction unit uses its output data set as input and its output data set as the output label set.
  • its model output approximates the labeled data set. In this way, the model pairing of the information reconstruction unit to the information generation unit is completed.
  • the reverse is also similar.
  • the information reconstruction part can be trained first. During the training process, a model part paired with it can be used to facilitate training. After the training is completed, the information reconstruction part will share all or part of its input and output data sets with the information.
  • the generation part is used for training the information generation part so that the output of the information generation part approximates the input of the information reconstruction part.
  • the above process can be completed offline.
  • the above description of training first and last is just for convenience of description, and it can also be done on both sides at the same time.
  • the input and output formats of the information generation department and the information reconstruction department have been agreed or predefined, and the input and output formats can be aligned so that the two can be trained separately. After training, they can share all or part of the input and output data sets with each other to form aggregated data. set for retraining on both sides.
  • the input data set of the information generation part and the output data set of the information reconstruction part adopt a common data set (the common data set can be agreed upon by both parties in advance, or stored in advance, or the data set can be generated by a predefined method. , etc.).
  • the information generation part uses the common data set as the input data set for training.
  • the output data set is shared with the information reconstruction part as the input data set of the information reconstruction part.
  • the information reconstruction part uses the common data set.
  • the data set is a labeled data set, and the purpose of training is to make the output data set under the input data set approximate the labeled data set.
  • the pairing of the information reconstruction unit to the information generation unit is completed. The pairing process from the information generation unit to the information reconstruction unit is similar and will not be described again.
  • the data set sharing method of the above-mentioned information generation unit and information reconstruction unit is not limited and can be realized through remote network transmission or other methods.
  • the corresponding information generation part and information reconstruction part form a bilateral model.
  • the model After undergoing relevant RAN4 tests or other predefined tests, when the performance meets the requirements, the model can register the corresponding model ID and version ID on the network side.
  • the training data set used by it the common data set can be bound to the model ID.
  • suitable data can be obtained at a relatively low cost, thereby supporting the bilateral network model architecture.
  • the data set can also be uploaded to the network side.
  • the network side or the terminal side may encounter model mismatch problems during operation. For example, after a terminal device moves to a certain cell, the network side or the terminal side discovers that the bilateral model IDs corresponding to the two are different, or the IDs are the same but the version information is different. Pairing can be achieved via air interface or other means.
  • X and Y can be shared.
  • the sending device terminal device or network device
  • the sending device can generate all or part of the input data set ⁇ X (data configuration, data index
  • This data set can be carried over a control channel or a data channel.
  • the information reconstruction part of the receiving device uses ⁇ Y (data configuration, data index
  • the model retraining and updating of the information generation part unchanged and the information reconstruction part. If the performance of the model retraining meets the requirements, for example, after further testing with or without air interface, the information reconstruction part will The model that should be retrained and updated is marked with the model ID (including version identification).
  • Y can be shared.
  • the sending device (such as a terminal device or a network device) outputs the entire data set or part of the data of ⁇ Y (data configuration, data index
  • the set is sent to the receiving device (such as network device or terminal device) through the air interface.
  • the data set can be carried over the control channel or the data channel,
  • the information reconstruction part of the receiving device uses ⁇ Y (data configuration, data index
  • the receiving device can obtain them in a way that has been stored in the receiving The device is obtained through the model ID, or obtained from the remote server through the model ID, or the corresponding data is generated from the data index according to predefined rules.
  • the model retraining and updating of the information generation part and the information reconstruction part can be achieved unchanged. If the performance of the model retraining meets the requirements, the information reconstruction part can be retrained after further testing (or without).
  • the model that should be retrained and updated will be marked as the model ID (including version identification).
  • the sending device (such as a terminal device or a network device) sends the model ID (including version) corresponding to the information generation unit and/or the input data configuration information corresponding to the model to the receiving device (such as a network device or a terminal device) through the air interface. ).
  • the model ID and/or the model corresponding input data configuration information can be carried through the control channel or the data channel. If the model ID of the information reconstruction part of the receiving device is inconsistent with the model ID corresponding to the information generation part, the receiving device can obtain the ⁇ Y(data configuration, Data index
  • data) ⁇ corresponding to the model ID of the information generation unit is obtained from the memory or downloaded from the remote server, and ⁇ X (data configuration, data index
  • the index generates corresponding data.
  • the second device sends an AI/ML related capability query to the first device; and the second device receives an AI/ML related capability response fed back by the first device.
  • the AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, data configuration information, AI/ML support training capability information, AI/ML support information ML upgrade capability information.
  • the model identification and/or version information of the information generation unit is different from the model identification and/or version information of the information reconstruction unit.
  • the version information is different, the second device acquires the first data and/or the second data corresponding to the model identification and/or version information of the information generation unit of the first device.
  • the second device sets the model identification and/or version information of the information reconstruction unit. It is the same as the model identification and/or version information of the information generation unit.
  • the second device sends confirmation information to the first device confirming that its information generation part can be used, and/or sends a message to enable the information generation part and/or the information reconstruction part. instructions.
  • the network device inquires the terminal device about AI/ML related capabilities.
  • the inquiry may include one or more of the following: signal processing module information, AI/ML support information , AI/ML model identification information, version information, AI/ML model data configuration information, etc.
  • the terminal device performs corresponding response reporting.
  • the network device when the network device discovers that the terminal device reports that a certain module has AI/ML capabilities, but the model ID of the information generation part it has is different from the model ID of the information reconstruction part of the network device, the network device can configure the terminal device to report:
  • the predefined X1 data set or the preconfigured first data set is input as the information generation unit, the data configuration information of the corresponding output Y1 data set, such as size, structure and other information.
  • the format of the above information reporting is predefined.
  • the network device when the network device discovers that the terminal device reports that a corresponding module has AI/ML capabilities, and the model ID of the information generation part it has is the same as the model ID of the information reconstruction part of the network device but the version is different.
  • Network device configuration terminal device reporting When the predefined X11 data set or the preconfigured first data set is input as the generation unit, the data configuration information of the corresponding output Y11 data set, for example, includes data set size, data set structure information (such as multiple data set blocks) and other information. The format of the above information reporting is predefined.
  • X11 and Y11 are parts of X1 and Y1 respectively.
  • the network device configures or schedules the terminal device to report the Y1 data set or Y11 data set according to the data configuration information reported by the terminal device.
  • the Y1 data set or Y11 data set is included in the uplink control information (UCI) or reported by the AI/ML data auxiliary information (via RRC message).
  • the network device After receiving the Y1 or Y11 data set reported by the terminal device, the network device uses the Y1 or Y11 data set as the input data of its information reconstruction unit.
  • the network device stores the X1 or X11 data set corresponding to the Y1 or Y11 data set, or the network device can download the corresponding X1 or X11 data set from the remote server.
  • the information reconstruction part of the network device uses the X1 or X11 data set as the label data set, thus retraining the information reconstruction part.
  • the network device stores the trained neural network model and parameters, and uses the model ID and version information of the information generation unit as signs of the neural network model and parameters. Then, the network device notifies the terminal device to enable the information generation part of the model ID and version information.
  • the network device can configure or schedule the terminal device to report the Y2 data set or Y21 data set.
  • the Y21 data set is part of the Y2 data set.
  • the training of the reconstruction part ends.
  • the network device stores the trained neural network model and parameters, and uses the model ID and version information of the information generation unit as signs of the neural network model and parameters.
  • the network device notifies the terminal device to enable the information generation part of the model ID and version information.
  • the network device informs the terminal device that the model ID and version cannot be used.
  • the network device inquires the terminal device about AI/ML related capabilities.
  • the inquiry may include one or more: signal processing module information, AI/ML support information, AI/ML model identification information, version information, AI/ML model data configuration information, etc.
  • the terminal device performs corresponding response reporting.
  • the network device uses the information Generate the model ID and version information of the department, and search the information corresponding to the model ID and version in its memory or the remote server.
  • the network device is trained based on the input and output data set. When, after training, the performance of the information reconstruction unit reaches the corresponding performance requirements, the training of the information reconstruction unit ends.
  • the network device stores the trained neural network model and parameters, and uses the model ID and version information of the information generation unit as signs of the neural network model and parameters.
  • the network device can notify the terminal device to enable the information generation part of the model ID and version information.
  • the second device acquires the first data X that is input to the information generation unit; and the second device acquires the second data Y corresponding to the first data X and output from the information generation unit.
  • suitable data can be obtained at a relatively low cost, thereby supporting the bilateral network model architecture.
  • the embodiment of the present application provides a data acquisition method, which is explained from the first device side, and the same content as the embodiment of the first aspect will not be described again.
  • Figure 12 is a schematic diagram of the data acquisition method according to the embodiment of the present application. As shown in Figure 12, the method includes:
  • the first device obtains the first data X output by the information reconstruction unit;
  • the first device acquires the second data Y that is input to the information reconstruction part and corresponds to the first data X.
  • the first device inputs the first data X into the information generation part, and uses the second data Y as label data to train the information generation part.
  • the second data Y is data input to the information reconstruction part and corresponds to the first data X; the first data The second data Y is input to the information reconstruction unit and generated.
  • the information generation part in the first device and the information reconstruction part in the second device use the first data X and the second data Y to perform training respectively;
  • the output is the second data Y or data that is similar to the second data Y;
  • the information generation unit that has been trained in the second device When the second data Y is input, the information reconstruction unit outputs the first data X or data approximate to the first data X.
  • the first data X and the second data Y are paired data sets; the paired data sets have model identification information.
  • one first data X corresponds to multiple second data Y.
  • the first device obtains the second data and/or the third data from inside the first device or from outside the first device according to identification information related to the information reconstruction part.
  • One data One data.
  • the identification information related to the information reconstruction part includes: the model identification and/or version information corresponding to the information reconstruction part, and/or the data of the model corresponding to the information reconstruction part Configuration information.
  • the second data Y is carried through a control channel or a data channel and transmitted over an air interface, or the second data Y is generated through a data index according to predefined rules, or the index of the second data Y is determined by The second device sends the information to the first device through the air interface.
  • the first data X is pre-stored in the first device, or the first data X is generated through a data index according to predefined rules, or the first data
  • the second device sends it to the first device through the air interface, or the index of the first data X is sent by the second device to the first device through the air interface.
  • the AI/ML model has a model identification and a version identification; the information generation part and the information reconstruction part of the same AI/ML model use the same model identification and version identification, and the information generation part and the information reconstruction part use the same model identification and version identification.
  • the Information Reconstruction Department has different sub-identities.
  • the sending device (such as a network device or a terminal device) sends the input ⁇ Y (data configuration, data index
  • This data set can be carried over a control channel or a data channel.
  • the information generation part of the receiving device uses ⁇ X (data configuration, data index
  • the purpose of training is to make x (data configuration, data index
  • the output of the generating part is equal to or approximately equal to y(data configuration, data index
  • the receiving device can obtain it by storing it in the receiving device and obtaining it through the model ID, or obtaining it from the remote server through the model ID, or according to predefined rules Corresponding data is generated from the data index.
  • the information generation part will retrain and update the model accordingly.
  • the trained and updated models are marked with the model ID (including version identification).
  • the first device sends an AI/ML related capability query to the second device; and the first device receives an AI/ML related capability response fed back by the second device.
  • the AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, data configuration information, AI/ML support training capability information, AI/ML support information ML upgrade capability information.
  • the model identification and/or version information of the information generation unit is different from the model identification and/or version information of the information reconstruction unit.
  • the version information is different, the first device acquires the first data and/or the second data corresponding to the model identification and/or version information of the information reconstruction unit of the second device.
  • the first device sets the model identification and/or version information of the information generation unit to be the same as The model identification and/or version information of the information reconstruction part are the same.
  • the first device sends confirmation information to the second device confirming that its information reconstruction part can be used, and/or enables the information generation part and/or the information reconstruction part instructions.
  • the first device obtains the first data X output by the information reconstruction unit; and the first device obtains the second data input to the information reconstruction unit and corresponding to the first data X Y.
  • suitable data can be obtained at a relatively low cost, thereby supporting the bilateral network model architecture.
  • the embodiment of this application takes CSI generation and reporting as an example to explain from the network device side.
  • the AI/ML model includes a CSI generation unit located in the terminal device and a CSI reconstruction unit located in the network device. The same content as the embodiments of the first and second aspects will not be described again.
  • the shared paired data set uses a standard predefined codebook or a DFT matrix similar to the codebook structure to generate an approximate channel matrix or channel matrix eigenvector, which greatly reduces the data set. overhead, especially if the data set needs to be transferred directly or indirectly through the air interface.
  • Figure 13 is an example diagram of constructing a data set according to the embodiment of the present application, which exemplarily shows the use of codebooks to construct a data set corresponding to a channel feature vector.
  • the codebook can approximate a channel matrix eigenvector with high accuracy through the linear combination of DFT vectors.
  • the codebook matrix can approximate the eigenvector that needs feedback. , or eigenvector matrix.
  • the codebook vector part can be generated on each side through a simple indexing method. In this way, only the bit vectors compressed by the CSI generation part need to be shared in the bilateral model part, which greatly reduces the overhead.
  • Table 5.2.2.2.1-2 Supported configurations of(N 1 ,N 2 )and(O 1 ,O 2 )
  • Table 2 schematically shows the configuration information of different port numbers of CSI-RS, corresponding to the configuration information of the codebook.
  • Table 5.2.2.2.3-5 Codebook for 1-layer and 2-layer CSI reporting using antenna ports 3000to2999+PCSI-RS
  • Table 3 schematically illustrates the PMI of codebook type2.
  • PMI i.e., i 1, ,q 1 ,q 2, n 1, n 2, ... such indexes
  • a codebook vector or codebook matrix can be constructed.
  • Other types of code The principles of the book are similar, and you can refer to related technologies without further explanation here.
  • codebook or codebook vector has been schematically described above, and the present application is not limited thereto.
  • Figure 14 is a schematic diagram of the data acquisition method according to the embodiment of the present application. As shown in Figure 14, the method includes:
  • the network device obtains the specific codebook vector input to the CSI generation unit.
  • the network device obtains the CSI bit information output from the CSI generation unit corresponding to the codebook vector.
  • the AI/ML model is a bilateral model with a model identification and a version identification; the CSI generation part and the CSI reconstruction part of the one bilateral model have the same model identification and version identification, and the The CSI generation unit and the CSI reconstruction unit have different sub-identities.
  • the terminal device including the CSI generation unit AE-x and the network device including the CSI reconstruction unit DE-y have not undergone joint training.
  • the network device inputs the CSI bit information into the CSI reconstruction part, and uses the codebook vector as label data to train the CSI reconstruction part.
  • the codebook vector (X) is data input to the CSI generation unit; the CSI bit information (Y) is the codebook vector (X) input to the CSI generation unit. The bit sequence generated afterwards.
  • codebook vector can be shown in Table 4:
  • Table 5.2.2.2.5-5 Codebook for 1-layer.2-layer,3-layer and 4-layer CSI reporting using antenna ports3000to 2999+P CSI-RS
  • Table 4 is the eType codebook generation table. Based on this table, codebook vectors or codebook matrices can also be generated through each subscript index of W. For the specific content of various NR codebooks, please refer to 3GPP 38.214.
  • the CSI bit information is carried by a control channel or a data channel and transmitted via an air interface, and/or the codebook vector is generated through a data index according to predefined rules.
  • a collection of bit sequences. (i1,i2,...,in) is the index (index) of the corresponding codebook, or it is called PMI, or it is called an indicator sequence. It is specifically based on the definitions of different types of codebooks. This is just a schematic explanation.
  • FIG. 15 is an example diagram of data acquired by the CSI reconstruction unit according to the embodiment of the present application.
  • the terminal device including the CSI generation unit AE-x inputs the sequence set corresponding to the sequence set ⁇ V AE-x (x1, x2,...,xn) ⁇ or the PMI set ⁇ x1, x2,...,xn ⁇ and
  • the corresponding output sequence ⁇ C AE-x (x1, x2, .., xn) ⁇ is sent to the network device including the CSI reconstruction unit DE-y.
  • DE-y takes ⁇ C AE-x (x1,x2,..,xn) ⁇ as its input sequence set, since it has received the codebook sequence set or the PMI set ⁇ x1,x2,..,xn ⁇ , DE- y can generate the corresponding codebook vector set ⁇ V AE-x (x1,x2,...,xn) ⁇ according to the generation rules of the corresponding codebook of NR.
  • Figure 16 is an example diagram of training performed by the CSI reconstruction unit according to the embodiment of the present application.
  • DE-y uses ⁇ C AE-x (x1,x2,..,xn) ⁇ and ⁇ V AE-x (i1,i2,...,in) ⁇ as data sets to train its AI/ML model.
  • the trained The purpose is to make the output vector approximate V AE-x (i1,i2,...,in) when the input is C AE -x (x1,x2,...,xn).
  • V AE-x i1,i2,...,in
  • the pairing process of AE-x and DE-y is completed.
  • the network device obtains the CSI bit information and/or the codebook vector from inside the network device or from outside the network device according to identification information related to the CSI generation unit.
  • the identification information related to the CSI generation unit includes: the model identification and/or version information corresponding to the CSI generation unit, and/or the data configuration information of the model corresponding to the CSI generation unit, And/or, the data index of the CSI generating unit, and/or the indication sequence of the codebook vector or DFT vector of the CSI generating unit.
  • the network device finds that the model ID or version of its CSI reconstruction unit does not match the model ID or version of the CSI generation unit of the terminal device, it obtains the corresponding input and output data based on the model ID or version information of the CSI generation unit. set.
  • the data set corresponding to the model ID and version can be obtained through internal storage or through a remote server.
  • the data set is, for example, a codebook sequence set ⁇ x1, x2,...,xn ⁇ and an output set ⁇ C AE-x (x1, x2,...,xn) ⁇ of the CSI generation unit. Because both are bit sequences, the amount of storage is greatly reduced.
  • the corresponding generation rule can generate the codebook vector ⁇ V AE-x (x1,x2,...,xn) ⁇ . Furthermore, the CSI reconstruction unit can be trained. Then use the model ID and version information of the CSI generation part to mark the corresponding model parameters when the training of the CSI reconstruction part is completed.
  • Figure 17 is another example diagram of data obtained by the CSI reconstruction unit according to the embodiment of the present application.
  • Codebooks based on standard definitions, such as Type 1, Type2, etype 2 codebook and other NR-defined codebooks, can use the precoding vector set ⁇ V AE-x (i1,i2,...,in) ⁇ generated by the codebook as CSI
  • the input of the AI/ML model of the generation part AE-x (CSI generation part), ⁇ C AE-x (i1,i2,..,in) ⁇ is the corresponding output bit sequence set.
  • (i1,i2,...,in) is the index sequence, or indicator sequence, or PMI of the corresponding codebook, and its arrangement order is predefined or preconfigured, or configured.
  • the terminal device including the CSI generation unit AE-x sends the output sequence ⁇ C AE-x (i1,i2,..,in) ⁇ to the network device including the CSI reconstruction unit DE-y.
  • Figure 18 is another example diagram of training performed by the CSI reconstruction unit according to the embodiment of the present application.
  • Figure 19 is another example diagram of training performed by the CSI reconstruction unit according to the embodiment of the present application.
  • DE-y takes ⁇ C AE-x (i1,i2,..,in) ⁇ as its input sequence set, since the order of the codebook index or flags (i1,i2,..,in) is known, DE-y Accordingly, the corresponding codebook vector set ⁇ V AE-x (i1,i2,...,in) ⁇ can be generated according to the generation rules of the corresponding codebook of NR.
  • DE-y uses ⁇ C AE-x (i1,i2,..,in) ⁇ and ⁇ V AE-x (i1,i2,...,in) ⁇ as data sets to train its AI/ML model. The trained The purpose is to make the output vector approximate V AE-x (i1,i2,...,in) when the input is C AE-x (i1,i2,...,in). When training is completed, the pairing process of AE-x and DE-y is completed.
  • the network device sends model training instructions and configuration information for model identification to the terminal device;
  • the configuration information includes an instruction sequence of a training codebook vector or a DFT matrix vector;
  • the network device receives An indication sequence sent by the terminal device for indicating a codebook vector or a DFT vector and a CSI generation unit output bit sequence corresponding to the vector.
  • the network device generates a corresponding codebook vector or DFT matrix vector based on the indication sequence according to predefined rules.
  • the network device receives a resource request from the terminal device; and the network device configures uplink resources for the terminal device, so that the terminal device sends the indication sequence and the CSI generation unit output bit sequence.
  • the network device sends an AI/ML related capability query to the terminal device; and the network device receives an AI/ML related capability response fed back by the terminal device.
  • the AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, data configuration information, AI/ML training capability information, AI/ML Upgrade capability information.
  • the model identification and/or version information of the CSI generation unit is different from the model identification and/or version information of the CSI reconstruction unit.
  • the terminal device is instructed or configured to send the training CSI bit sequence and/or the corresponding codebook vector corresponding to the model identification and/or version information of the CSI generation unit.
  • the network device sets the model identification and/or version information of the CSI reconstruction unit to The same as the model identification and/or version information of the CSI generation unit.
  • the network device sends instruction information to enable the CSI generation unit and/or the CSI reconstruction unit to the terminal device, and/or sends confirmation that the CSI generation unit can use Confirm the message.
  • Figure 20 is a signaling diagram of CSI reconstruction unit training according to an embodiment of the present application.
  • the network side allows the terminal side to report the AI/ML capabilities of the CSI feedback, which can include inquiries about AI/ML model ID, version and other model information.
  • the terminal side reports relevant capability information, which includes instructions for the AI/ML model ID, version and other model information used by its CSI generation unit. If the network side has the CSI reconstruction department model with the corresponding ID and version, it enters the joint working phase.
  • the network side can configure the terminal side to report corresponding data and train the CSI reconstruction unit.
  • the network side can query the terminal side for format information related to the bit sequence output by the CSI generation unit, and the terminal side reports the output bit information of its CSI generation unit according to a predefined format, such as the number of output information bits corresponding to each codebook vector or matrix. .
  • the network side can instruct the terminal side to send the data configuration information of the CSI output set corresponding to the corresponding codebook vector set (for example, the bit length corresponding to each codebook vector), and send the CSI output set reporting configuration.
  • the terminal side can report the corresponding CSI output set according to the configuration/instruction.
  • the network side uses the codebook vector corresponding to the codebook serial number as label data, uses the CSI generator output corresponding to the codebook serial number as input data, and trains the model of the CSI reconstruction part so that the model output is close to the label data.
  • the model training is completed, update the model and update the model flag information.
  • the network side can configure and instruct the terminal side to use the CSI generation unit to perform CSI measurement and feedback; the terminal side uses the CSI generation unit to generate CSI for the measurement results according to the configured CSI measurement resources, and reports CSI according to the configured resources. .
  • the network side may use the updated CSI reconstruction unit to reconstruct the received CSI.
  • the CSI generation unit may use the channel matrix after channel estimation as input, or the eigenvector of the channel matrix after channel estimation as input.
  • the number of input feature vectors can be selected based on rank, or based on the feature vectors of a single layer.
  • the CSI generation department or CSI reconstruction department needs to communicate the input and output possibilities of the CSI generation department or CSI reconstruction department when acquiring the data set and performing retraining and fine tuning.
  • Configuration information such as input data length, input data dimensions, etc.
  • the format and/or size of the CSI is predefined, or the format and/or size of the CSI is configured by the network device.
  • the network side needs to know through inquiry that the CSI generation unit can generate multiple CSI bit lengths.
  • the network side may configure the terminal side to feedback the output of the CSI generation unit of a certain bit length, or multiple bit lengths, or all bit lengths.
  • the output of the CSI generation unit may be configured to know through inquiry that the CSI generation unit can generate multiple CSI bit lengths.
  • Table 5 is an example of information related to the CSI generation unit that needs to be fed back by the terminal side.
  • Codebook indicator bit (PMI) code book vector CSI generation unit bit 1 CSI generation unit bit 2 CSI generation unit bit 3 A1 A2 A3 A4 A5 B1 B2 B3 B4 B5
  • the network device expands the codebook vector and uses the expanded codebook vector as the label data.
  • the codebook vectors generated may not be enough to complete the training of AI/ML networks, and more data sets are needed. train. Or, it is necessary to select a codebook set that is more suitable for the CSI generation part and the CSI reconstruction part.
  • the number of optional beams of W1 is at most 6, 8, 10, 12, and 16; the mapping angle of W2 is at most 5 bits, and the amplitude is at most 5 bits.
  • the existing codebook set can be extended to form an extended codebook set or a codebook set for CSI model training.
  • the existing W2 mapping angle is 2 or 3 bits
  • the amplitude is 3 bits, which can be extended to 4 bits for angle, 4 bits for amplitude, etc. M can also take larger numbers than currently supported by the standard.
  • the above codebook set for CSI model training can be configured to obtain the complete set of codebook vectors that can be generated by one of the codebooks defined by the current standard (Rel-15, 16, 17, 18, etc.), or according to a certain rule subset. Alternatively, you can obtain an expanded set of codebooks through configuration and select subsets based on rules. Or use other codebook index sequences to call the above possible codebook set. The method of calling the codebook sequence is known to both the CSI generation part and the CSI reconstruction part.
  • the network device finds that the model of the terminal device is inconsistent, either the CSI generation unit performs online training and upgrades to match the CSI reconstruction unit, or the CSI reconstruction unit performs online training and upgrades to match the CSI Generate partial matches.
  • the network side can communicate the configuration information of the CSI reconstruction unit with the terminal side.
  • the network side may schedule the terminal side to send the index of the codebook vector to be used by the CSI generation unit and the corresponding output of the CSI generation unit.
  • This call can be performed in batches; for example, a part of the data is called first, and if the training effect is not up to standard, another part of the data set is scheduled. Training is done through multiple calls.
  • all possible codebook vectors or all possible vectors corresponding to the DFT transformation matrix can be generated based on a predefined codebook or DFT transformation matrix (for example, a two-dimensional transformation matrix in the angle delay domain).
  • all possible codeword vectors are arranged in a predefined order.
  • the vectors corresponding to the DFT transformation matrix can also be predefined to generate all corresponding vectors. for example,
  • the specific DFT adopted is based on pre-definition. More W cascade methods can be used to form the basis functions of the spatial domain, frequency domain, and time domain feature mapping of the channel.
  • the above is a certain model ID and version, the complete set of CSI generation vectors and the complete set of possible outputs. Subsets can be obtained through the index, and the subsets are used for training. Multiple subsets of the full set can be called multiple times for training.
  • the call to a subset is indicated by an index or a bitmap of the index set.
  • the length of the bitmap corresponds to the length of all serial numbers in the table, one-to-one correspondence from left to right.
  • a 0 in the bitmap means using the vector, a 1 means not using the vector.
  • the index set here can correspond to different subbands, indicating the indexes corresponding to different subbands.
  • the terminal side or the network side can, through data screening and collection, extract the PMI corresponding to the corresponding typical channel vector to form a data set for the current channel or the channel experienced in history, and index the PMI according to the bitmap instructions.
  • the codebook vectors in the channel spatial domain and frequency domain correspond to each other.
  • the angle and time delay here are respectively corresponding to different DFT basis vectors. It is worth noting that the above is only a schematic explanation. Similarly, a data set corresponding to broadband CSI or a data set corresponding to channel characteristics in other ways can also be formed according to the configuration. These data sets can be bound to the model ID, or based on the network configuration, further bound based on the cell model ID, or bound based on the UE model ID, or identified based on other rules.
  • the shared paired data set can reflect the characteristics of the channel in the air domain, frequency domain, and time domain, its identification method can reflect these characteristics accordingly, making it easier to apply more information when the terminal or base station side can obtain the channel characteristic information.
  • a data set suitable for the current channel conditions for training Corresponding to the channel vector data set expressed in the above manner, a corresponding generating unit compressed information data set is also required. That is, all channel vector data sets with bitmap 1. This forms a paired data set.
  • the terminal side After obtaining the configuration of the network side, the terminal side sends the above bitmap set selected by its generation unit and the corresponding generation unit output set to the network side.
  • the data set sample is large, the air interface overhead is too large. If the terminal side and the network side can obtain a common paired data set, air interface interaction can be achieved by transmitting the data set ID over the air interface, so that the air interface overhead is small.
  • the base station equipment monitors the CSI feedback performance and determines that the model for a certain model ID (including version information), or the performance of the CSI model being used is not good, or the base station decides to start CSI reconstruction based on other judgments. model training.
  • the base station configures the terminal equipment to send the index flag information (such as bitmap information) related to the channel state information (CSI) of the CSI generation unit for the currently used CSI model ID or a certain model ID, as well as the relevant training codebook vector or DFT matrix vector. Instructions and configuration information.
  • index flag information such as bitmap information
  • CSI channel state information
  • the terminal device Based on the model ID of the currently used CSI generation unit or the model ID configured by the base station, and the above-mentioned vector indication configuration information, the terminal device sends to the base station the index flag information related to the channel state information (CSI) of the CSI generation unit corresponding to the model (such as bitmap information), or index flag information set.
  • CSI channel state information
  • the base station receives the index information or index information set, and obtains the input information set required by the CSI reconstruction unit of the model ID based on the model ID being used or its configured model ID. For the model ID (including version information) and based on the index or index set, the base station obtains the corresponding CSI generation unit output vector set from its internal or remote server, and generates the corresponding CSI generation unit input vector set based on the index or index set. .
  • the CSI generation unit output vector set is used as the input data of the CSI reconstruction unit, and the CSI generation unit input vector set is used as the label data of the CSI reconstruction unit to train the CSI reconstruction unit.
  • the base station can send the codebook matrix or DFT matrix indication and configuration information for a certain model ID and version, as well as the corresponding index information or index information set to the terminal device, and the terminal device downloads it from the local or remote server accordingly.
  • the corresponding CSI generation unit (or CSI reconstruction unit) inputs a vector set (output vector set, vector corresponding to the codebook or DFT transformation matrix) and output vector set (input vector set).
  • the CSI generation part and the CSI reconstruction part have multiple data sets for different channel environments for training.
  • the data set is given as a codebook vector set, or a generating vector of a DFT transformation matrix (such as a two-dimensional transformation matrix in the angle delay domain), such as
  • the vectors of each sub-matrix are given based on the codebook definition or other predefined definitions. In this way, based on the model ID, version information, codebook or DFT matrix configuration information, and data set serial number, the base station and the terminal device can obtain the training data set required by the CSI reconstruction unit or CSI generation unit.
  • the base station can configure the terminal device to send the above-mentioned data set serial number of the CSI generation unit, or the CSI generation unit data set serial number for the model ID version information, and obtain the training data set required by the CSI reconstruction unit based on this information.
  • the base station can send training instructions, codebook or DFT matrix instructions and configuration information to the terminal equipment, as well as the data set serial number of the CSI reconstruction part, or the data set serial number of the CSI reconstruction part for model ID version information.
  • the terminal device obtains the data set for training, and sends training completion information to the base station after the training is completed.
  • its training completed output data set can be used to update the corresponding data set of the server.
  • the network device obtains a specific codebook vector input to the CSI generation unit; and the network device obtains the CSI bit information output from the CSI generation unit corresponding to the codebook vector.
  • suitable data can be obtained at a relatively low cost, thereby supporting the bilateral network model architecture.
  • the embodiment of this application takes CSI generation and reporting as an example to explain from the terminal device side.
  • the AI/ML model includes a CSI generation unit located in the terminal device and a CSI reconstruction unit located in the network device. The same content as the embodiments of the first to third aspects will not be described again.
  • Figure 21 is a schematic diagram of the data acquisition method according to the embodiment of the present application. As shown in Figure 21, the method includes:
  • the terminal device obtains the codebook vector output by the CSI reconstruction unit.
  • the terminal device obtains the CSI bit information input to the CSI reconstruction unit and corresponding to the codebook vector.
  • the AI/ML model is a bilateral model with a model identification and a version identification; the CSI generation part and the CSI reconstruction part of the same AI/ML model have the same model identification and version identification, and the CSI generation The CSI reconstruction part and the CSI reconstruction part have different sub-identities.
  • the CSI bit information (Y) is information input to the CSI reconstruction unit and corresponds to the codebook vector; the codebook vector (X) is part or all of the data of a specific data set , generated after the CSI bit information (Y) is input to the CSI reconstruction unit.
  • the CSI bit information is carried by a control channel or a data channel and transmitted via an air interface, and/or the codebook vector is generated through a data index according to predefined rules.
  • the terminal device obtains the CSI bit information and/or the code from a memory or from inside the terminal device or from outside the terminal device according to the identification information related to the CSI reconstruction part. Book vector.
  • the identification information related to the CSI reconstruction unit includes: model identification and/or version information corresponding to the CSI reconstruction unit, and/or data of the model corresponding to the reconstruction generation unit. Configuration information, and/or, the data index of the CSI reconstruction part, and/or the indication sequence of the codebook vector or DFT vector of the CSI reconstruction part.
  • the terminal device receives the model training instructions and configuration information for the model identification sent by the network device; the configuration information includes an instruction sequence of a training codebook vector or a DFT matrix vector and the vector corresponding
  • the CSI reconstruction unit inputs the bit sequence.
  • the terminal device generates a corresponding codebook vector or DFT matrix vector according to predefined rules based on the indication sequence.
  • the terminal device receives a resource configuration of the network device; and the terminal device receives the indication sequence and the CSI reconstruction unit input bit sequence according to the resource configuration.
  • FIG. 22 is an example diagram of data obtained by the CSI generation unit according to the embodiment of the present application.
  • the network side can send the CSI data set corresponding to the codebook vector to the terminal device.
  • data or other acquisition methods please refer to the previous embodiments.
  • the terminal device inputs the codebook vector into the CSI generation part, and uses the CSI bit information as label data to train the CSI generation part.
  • Figure 23 is an example diagram of training performed by the CSI generation unit according to the embodiment of the present application.
  • the codebook vector set corresponding to the codebook vector sequence number can be used as input, and the CSI data set corresponding to the codebook vector set can be used as label data to train the CSI generation unit.
  • the terminal device receives an AI/ML related capability query sent by the network device; and the terminal device feeds back an AI/ML related capability response to the network device.
  • the AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, data configuration information, AI/ML training capability information, AI/ML Upgrade capability information.
  • the model identification and/or version information of the CSI generation unit is different from the model identification and/or version information of the CSI reconstruction unit.
  • the terminal device is instructed or configured to receive CSI bit information and/or codebook vectors corresponding to the model identification and/or version information of the CSI reconstruction unit.
  • the terminal device sets the model identification and/or version information of the CSI generation unit to be consistent with the The model identification and/or version information of the CSI reconstruction part are the same.
  • the terminal device sends confirmation information to the network device confirming that its CSI generation unit can be used, and/or indication information indicating that the CSI generation unit is updated.
  • Figure 24 is a signaling schematic diagram of training by the CSI generation unit according to the embodiment of the present application.
  • the network side allows the terminal side to report the AI/ML capabilities of the CSI feedback, which can include inquiries about model information such as AI/ML model ID and version.
  • the terminal side reports relevant capability information, which includes instructions for the AI/ML model ID, version and other model information used by its CSI generation unit. If the network side has the CSI reconstruction department model with the corresponding ID and version, it enters the joint working phase.
  • the network side knows that the AI/ML model ID and version used by the CSI generation unit on the terminal side do not match one or more CSI reconstruction units on the network side, it means that the AI/ML models on both sides of the communication cannot work together. .
  • the network side can instruct the terminal side to perform training upgrade and train the CSI generation unit.
  • the terminal side can send a training upgrade response.
  • the network side may form a data set based on the input information corresponding to the output codebook vector of the CSI reconstruction unit.
  • the data set may also be pre-stored on the network side.
  • the network side can send codebook related indication information, CSI data set corresponding to the codebook sequence, model identification information, etc. to the terminal side.
  • the terminal side uses the codebook vector corresponding to the codebook serial number as input, and uses the CSI data corresponding to the codebook serial number received from the network side as label data to train the CSI generation part, so that the model output approximates label data.
  • the model training is completed, the model and model identification information are updated. You can then indicate that the model update is complete and model pairing is complete.
  • the network side can configure and instruct the terminal side to use the CSI generation unit to perform CSI measurement and feedback; the terminal side uses the updated CSI generation unit to generate CSI based on the configured CSI measurement resources for the measurement results, and generates CSI according to the configured resources. Conduct CSI reporting. The network side may use the CSI reconstruction unit to reconstruct the received CSI.
  • the format and/or size of the CSI is predefined, or the format and/or size of the CSI is configured by the network device.
  • the terminal device expands the codebook vector and uses the expanded codebook vector as the label data.
  • the number of optional beams of W1 is at most 6, 8, 10, 12, and 16; the mapping angle of W2 is at most 5 bits, and the amplitude is at most 5 bits.
  • the terminal device receives the model training instruction information sent by the network device and/or the CSI reconstruction unit model data set instruction information.
  • the terminal device obtains the corresponding data set (from a local or remote server) as an output data set for training of the CSI generation unit, and obtains or generates an input data set of the CSI generation unit according to the data set serial number.
  • the above training instruction information may include model ID information, version information, codebook matrix or DFT transformation matrix instruction and configuration information, and data set instruction information, indicating the data subset corresponding to the model ID and version.
  • the terminal device can train the CSI generation unit based on these data.
  • the terminal device sends the training data set and continues to request information to the network device.
  • the network device continues to send the above CSI reconstruction unit model data set indication information to indicate other data sets.
  • the terminal device obtains the corresponding data set (from a local or remote server) as an output data set for training of the CSI generation unit, and obtains or generates an input data set of the CSI generation unit according to the data set serial number. The above process can continue until the model training and test performance meets the requirements, and the terminal device sends training completion indication information to the network device.
  • the network device can send the codebook matrix or DFT matrix indication and configuration information, and the corresponding index information or index information set to the terminal device for a certain model ID and version, and the terminal device accordingly obtains the codebook matrix or DFT matrix indication and configuration information from the local or slave device.
  • the remote server downloads the corresponding CSI generation unit (CSI reconstruction unit) input vector set (output vector set, codebook or DFT transformation matrix corresponding vector) and output vector set (input vector set).
  • the terminal device can train the CSI generation unit based on these data.
  • the terminal device sends the training data set to continue requesting information to the network device, and the network device sends other codebook matrix or DFT matrix indication and configuration information, and corresponding index information or index information set.
  • the terminal device continues training after obtaining the corresponding data set until the model training test performance meets the requirements, and the terminal device sends training completion instruction information to the network device.
  • the terminal device obtains the codebook vector output by the CSI reconstruction unit; and the terminal equipment obtains the CSI bit information input to the CSI reconstruction unit and corresponding to the codebook vector.
  • suitable data can be obtained at a relatively low cost, thereby supporting the bilateral network model architecture.
  • An embodiment of the present application provides a data acquisition device.
  • the device may be, for example, the aforementioned second device (terminal device or network device), or may be some or some components or components configured in the second device.
  • terminal device or network device may be some or some components or components configured in the second device.
  • FIG 25 is a schematic diagram of the data acquisition device according to the embodiment of the present application.
  • the AI/ML model includes an information generation part located in the first device and an information reconstruction part located in the second device.
  • the data acquisition device 2500 includes:
  • a first acquisition unit 2501 that acquires the first data X input to the information generation unit;
  • the second acquisition unit 2502 acquires the second data Y corresponding to the first data X and output from the information generation unit.
  • the data acquisition device 2500 may also include:
  • the first training unit 2503 inputs the second data Y into the information reconstruction part, and uses the first data X as label data to train the information reconstruction part.
  • the first data Corresponding data generated later.
  • the information generation part in the first device and the information reconstruction part in the second device use the first data X and the second data Y to perform training respectively;
  • the output is the second data Y or data that is similar to the second data Y; the information generation unit that has been trained in the second device
  • the information reconstruction unit outputs the first data X or data approximate to the first data X.
  • the first data X and the second data Y are paired data sets; the paired data sets have model identification information.
  • one first data X corresponds to multiple second data Y.
  • the first obtaining unit 2501 or the second obtaining unit 2502 obtains the information from inside the second device or from outside the second device according to the identification information related to the information generation part.
  • the identification information related to the information generation part includes: the model identification and/or version information corresponding to the information generation part, and/or the data configuration information of the model corresponding to the information generation part.
  • the second data Y is carried through a control channel or a data channel and transmitted over an air interface, or the second data Y is generated through a data index according to predefined rules, or the index of the second data Y is determined by The first device sends the information to the second device through the air interface.
  • the first data X is pre-stored in the second device, or the first data X is generated through a data index according to predefined rules, or the first data
  • the first device sends it to the second device through the air interface, or the index of the first data X is sent by the first device to the second device through the air interface.
  • the AI/ML model has a model identification and a version identification; the information generation part and the information reconstruction part of the same AI/ML model use the same model identification and version identification, and the information generation part and the information reconstruction part use the same model identification and version identification.
  • the Information Reconstruction Department has different sub-identities.
  • the device further includes:
  • a sending unit that sends an AI/ML related capability query to the first device
  • a receiving unit that receives the AI/ML related capability response fed back by the first device.
  • the AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, data configuration information, AI/ML support training capabilities Information, AI/ML upgrade capability information.
  • the model identification and/or version information of the information generation unit is different from the model identification and/or version information of the information reconstruction unit.
  • the first acquisition unit or the second acquisition unit acquires the first data and/or the third data corresponding to the model identification and/or version information of the information generation unit of the first device. 2 data.
  • the model identification and/or version information of the information reconstruction unit is set to be consistent with the information The model identification and/or version information of the generation part are the same.
  • the sending unit also sends confirmation information to the first device confirming that its information generation part can be used, and/or sends a message to enable the information generation part and/or the information reconstruction part. instructions.
  • the second device is a network device, and the AI/ML model includes a CSI generation unit located in the terminal device and a CSI reconstruction unit located in the network device.
  • the first acquisition unit 2501 acquires a specific codebook vector input to the CSI generation unit; and the second acquisition unit 2502 acquires the CSI output from the CSI generation unit corresponding to the codebook vector. bit information.
  • the first training unit 2503 inputs the CSI bit information into the CSI reconstruction part, and uses the codebook vector as label data to train the CSI reconstruction part.
  • the codebook vector (X) is data input to the CSI generation unit; the CSI bit information (Y) is the codebook vector (X) input to the CSI generation unit. The bit sequence generated afterwards.
  • the CSI bit information is carried by a control channel or a data channel and transmitted via an air interface, and/or the codebook vector is generated through a data index according to predefined rules.
  • the network device obtains the CSI bit information and/or the codebook vector from inside the network device or from outside the network device according to identification information related to the CSI generation unit.
  • the identification information related to the CSI generation unit includes: the model identification and/or version information corresponding to the CSI generation unit, and/or the data configuration information of the model corresponding to the CSI generation unit, And/or, the data index of the CSI generating unit, and/or the indication sequence of the codebook vector or DFT vector of the CSI generating unit.
  • the network device sends model training instructions and configuration information for model identification to the terminal device;
  • the configuration information includes an instruction sequence of a training codebook vector or a DFT matrix vector;
  • the network device receives an indication sequence sent by the terminal device for indicating a codebook vector or a DFT vector and a CSI generation unit output bit sequence corresponding to the vector.
  • the network device generates a corresponding codebook vector or DFT matrix vector based on the indication sequence according to predefined rules.
  • the network device receives a resource request from the terminal device; and the network device configures uplink resources for the terminal device, so that the terminal device sends the indication sequence and the CSI generation unit output bit sequence.
  • the AI/ML model is a bilateral model with a model identification and a version identification; the CSI generation part and the CSI reconstruction part of the one bilateral model have the same model identification and version identification, and the The CSI generation unit and the CSI reconstruction unit have different sub-identities.
  • the network device sends an AI/ML related capability query to the terminal device; and the network device receives an AI/ML related capability response fed back by the terminal device.
  • the AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, data configuration information, AI/ML training capability information , AI/ML upgrade capability information.
  • the model identification and/or version information of the CSI generation unit is different from the model identification and/or version information of the CSI reconstruction unit.
  • the terminal device is instructed or configured to send the training CSI bit sequence and/or the corresponding codebook vector corresponding to the model identification and/or version information of the CSI generation unit.
  • the format and/or size of the CSI is predefined, or the format and/or size of the CSI is configured by the network device.
  • the network device sets the model identification and/or version information of the CSI reconstruction unit to The same as the model identification and/or version information of the CSI generation unit.
  • the network device sends instruction information to enable the CSI generation unit and/or the CSI reconstruction unit to the terminal device, and/or sends confirmation that the CSI generation unit can use Confirm the message.
  • the network device expands the codebook vector and uses the expanded codebook vector as the label data.
  • the number of optional beams of W1 is at most 6, 8, 10, 12, and 16; the mapping angle of W2 is at most 5 bits, and the amplitude is at most 5 bits.
  • the data acquisition device 2500 may also include other components or modules.
  • the specific contents of these components or modules please refer to related technologies.
  • FIG. 25 only illustrates the connection relationships or signal directions between various components or modules, but it should be clear to those skilled in the art that various related technologies such as bus connections can be used.
  • Each of the above components or modules can be implemented by hardware facilities such as a processor, a memory, a transmitter, a receiver, etc.; the implementation of this application is not limited to this.
  • the second device acquires the first data X that is input to the information generation unit; and the second device acquires the second data Y corresponding to the first data X and output from the information generation unit.
  • appropriate data can be obtained at a relatively low cost to support the bilateral network model architecture.
  • An embodiment of the present application provides a data acquisition device.
  • the device may be, for example, the aforementioned first device (terminal device or network device), or may be some or some components or components configured in the first device.
  • the same contents as those in the first to fourth embodiments will no longer be included. Repeat.
  • FIG 26 is a schematic diagram of the data acquisition device according to the embodiment of the present application.
  • the AI/ML model includes an information generation unit located in the first device and an information reconstruction unit located in the second device.
  • the data acquisition device 2600 includes:
  • the third acquisition unit 2601 acquires the first data X output by the information reconstruction unit.
  • the fourth acquisition unit 2602 acquires the second data Y that is input to the information reconstruction unit and corresponds to the first data X.
  • the data acquisition device 2600 may also include:
  • the second training unit 2603 inputs the first data X into the information generation part, and uses the second data Y as label data to train the information generation part.
  • the second data Y is data input to the information reconstruction part and corresponds to the first data X; the first data The second data Y is input to the information reconstruction unit and generated.
  • the information generation part in the first device and the information reconstruction part in the second device use the first data X and the second data Y to perform training respectively;
  • the output is the second data Y or data that is similar to the second data Y; the information generation unit that has been trained in the second device
  • the information reconstruction unit outputs the first data X or data approximate to the first data X.
  • the first data X and the second data Y are paired data sets; the paired data sets have model identification information.
  • one first data X corresponds to multiple second data Y.
  • the first device obtains the second data and/or the third data from inside the first device or from outside the first device according to identification information related to the information reconstruction part.
  • One data One data.
  • the identification information related to the information reconstruction part includes: the model identification and/or version information corresponding to the information reconstruction part, and/or the data of the model corresponding to the information reconstruction part Configuration information.
  • the second data Y is carried through a control channel or a data channel and transmitted over an air interface, or the second data Y is generated through a data index according to predefined rules, or the index of the second data Y is determined by The second device sends the information to the first device through the air interface.
  • the first data X is pre-stored in the first device, or the first data X is generated through a data index according to predefined rules, or the first data
  • the second device sends it to the first device through the air interface, or the index of the first data X is sent by the second device to the first device through the air interface.
  • the AI/ML model has a model identification and a version identification; the information generation part and the information reconstruction part of the same AI/ML model use the same model identification and version identification, and the information generation part and the information reconstruction part use the same model identification and version identification.
  • the Information Reconstruction Department has different sub-identities.
  • the device further includes:
  • a sending unit that sends an AI/ML related capability query to the second device
  • a receiving unit that receives the AI/ML related capability response fed back by the second device.
  • the AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, data configuration information, AI/ML support training capabilities Information, AI/ML upgrade capability information.
  • the model identification and/or version information of the information generation unit is different from the model identification and/or version information of the information reconstruction unit.
  • the version information is different, the first device acquires the first data and/or the second data corresponding to the model identification and/or version information of the information reconstruction unit of the second device.
  • the first device sets the model identification and/or version information of the information generation unit to be the same as The model identification and/or version information of the information reconstruction part are the same.
  • the sending unit sends confirmation information to the second device confirming that its information reconstruction part can be used, and/or enabling the information generation part and/or the information reconstruction part. Instructions.
  • the first device is a terminal device.
  • the AI/ML model includes a CSI generation unit located in the terminal device and a CSI reconstruction unit located in the network device.
  • the third acquisition unit 2601 acquires the codebook vector output by the CSI reconstruction unit; and the fourth acquisition unit 2602 acquires the CSI bit information input to the CSI reconstruction unit and corresponding to the codebook vector.
  • the second training unit 2603 inputs the codebook vector into the CSI generation part, and uses the CSI bit information as label data to train the CSI generation part.
  • the CSI bit information (Y) is information input to the CSI reconstruction unit and corresponds to the codebook vector; the codebook vector (X) is part or all of the data of a specific data set , generated after the CSI bit information (Y) is input to the CSI reconstruction unit.
  • the CSI bit information is carried by a control channel or a data channel and transmitted via an air interface, and/or the codebook vector is generated through a data index according to predefined rules.
  • the terminal device obtains the CSI bit information and/or the code from a memory or from inside the terminal device or from outside the terminal device according to the identification information related to the CSI reconstruction part. Book vector.
  • the identification information related to the CSI reconstruction unit includes: model identification and/or version information corresponding to the CSI reconstruction unit, and/or data of the model corresponding to the reconstruction generation unit. Configuration information, and/or, the data index of the CSI reconstruction part, and/or the indication sequence of the codebook vector or DFT vector of the CSI reconstruction part.
  • the terminal device receives the model training instructions and configuration information for the model identification sent by the network device; the configuration information includes an instruction sequence of a training codebook vector or a DFT matrix vector and the vector corresponding
  • the CSI reconstruction unit inputs the bit sequence.
  • the terminal device generates a corresponding codebook vector or DFT matrix vector according to predefined rules based on the instruction sequence.
  • the terminal device receives a resource configuration of the network device; and the terminal device receives the indication sequence and the CSI reconstruction unit input bit sequence according to the resource configuration.
  • the AI/ML model is a bilateral model with a model identification and a version identification; the CSI generation part and the CSI reconstruction part of the same AI/ML model have the same model identification and version identification, and the CSI generation The CSI reconstruction part and the CSI reconstruction part have different sub-identities.
  • the terminal device receives an AI/ML related capability query sent by the network device; and the terminal device feeds back an AI/ML related capability response to the network device.
  • the AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, data configuration information, AI/ML training capability information , AI/ML upgrade capability information.
  • the model identification and/or version information of the CSI generation unit is different from the model identification and/or version information of the CSI reconstruction unit.
  • the terminal device is instructed or configured to receive CSI bit information and/or codebook vectors corresponding to the model identification and/or version information of the CSI reconstruction unit.
  • the format and/or size of the CSI is predefined, or the format and/or size of the CSI is configured by the network device.
  • the terminal device sets the model identification and/or version information of the CSI generation unit to be consistent with the The model identification and/or version information of the CSI reconstruction part are the same.
  • the terminal device sends confirmation information to the network device confirming that its CSI generation unit can be used, and/or indication information indicating that the CSI generation unit is updated.
  • the terminal device expands the codebook vector and uses the expanded codebook vector as the label data.
  • the number of optional beams of W1 is at most 6, 8, 10, 12, and 16; the mapping angle of W2 is at most 5 bits, and the amplitude is at most 5 bits.
  • the data acquisition device 2600 may also include other components or modules.
  • the specific contents of these components or modules please refer to related technologies.
  • FIG. 26 only illustrates the connection relationships or signal directions between various components or modules, but it should be clear to those skilled in the art that various related technologies such as bus connections can be used.
  • Each of the above components or modules can be implemented by hardware facilities such as a processor, a memory, a transmitter, a receiver, etc.; the implementation of this application is not limited to this.
  • the first device obtains the first data X output by the information reconstruction unit; and the first device obtains the second data input to the information reconstruction unit and corresponding to the first data X Y.
  • suitable data can be obtained at a relatively low cost, thereby supporting the bilateral network model architecture.
  • An embodiment of the present application also provides a communication system. Refer to FIG. 1 . Contents that are the same as those in the first to sixth embodiments will not be described again.
  • the AI/ML model includes an information generation part located in the first device and an information reconstruction part located in the second device.
  • the communication system 100 may at least include:
  • a first device that acquires the first data X output by the information reconstruction part; and acquires the second data Y input to the information reconstruction part and corresponding to the first data X;
  • a second device acquires the first data X input to the information generating section; and acquires the second data Y corresponding to the first data X and output from the information generating section.
  • the AI/ML model includes a CSI generation part located in the terminal device and a CSI reconstruction part located in the network device; the communication system 100 may at least include:
  • a network device that acquires a specific codebook vector input to the CSI generation unit; and acquires CSI bit information output from the CSI generation unit corresponding to the codebook vector;
  • a terminal device acquires a codebook vector output by the CSI reconstruction unit; and acquires CSI bit information input to the CSI reconstruction unit and corresponding to the codebook vector.
  • the embodiment of the present application also provides a network device, which may be a base station, for example, but the present application is not limited thereto and may also be other network devices.
  • a network device which may be a base station, for example, but the present application is not limited thereto and may also be other network devices.
  • FIG 27 is a schematic structural diagram of a network device according to an embodiment of the present application.
  • network device 2700 may include a processor 2710 (eg, a central processing unit CPU) and a memory 2720; the memory 2720 is coupled to the processor 2710.
  • the memory 2720 can store various data; in addition, it also stores an information processing program 2730, and the program 2730 is executed under the control of the processor 2710.
  • the processor 2710 may be configured to execute a program to implement the data acquisition method as described in the embodiment of the first aspect.
  • the processor 2710 may be configured to perform the following control: acquire the first data X input to the information generation unit; and acquire the second data Y corresponding to the first data X and output from the information generation unit.
  • the processor 2710 may be configured to execute a program to implement the data acquisition method as described in the embodiment of the second aspect.
  • the processor 2710 may be configured to perform the following control: acquire the first data X output by the information reconstruction part; and acquire the second data X input to the information reconstruction part and corresponding to the first data X. DataY.
  • the processor 2710 may be configured to execute a program to implement the data acquisition method as described in the embodiment of the third aspect.
  • the processor 2710 may be configured to perform the following control: acquire a specific codebook vector input to the CSI generation unit; and acquire CSI bit information output from the CSI generation unit corresponding to the codebook vector.
  • the network device 2700 may also include: a transceiver 2740, an antenna 2750, etc.; the functions of the above components are similar to those of the existing technology and will not be described again here. It is worth noting that the network device 2700 does not necessarily include all components shown in Figure 27; in addition, the network device 2700 may also include components not shown in Figure 27, and reference can be made to the existing technology.
  • the embodiment of the present application also provides a terminal device, but the present application is not limited to this and may also be other devices.
  • Figure 28 is a schematic diagram of a terminal device according to an embodiment of the present application.
  • the terminal device 2800 may include a processor 2810 and a memory 2820; the memory 2820 stores data and programs and is coupled to the processor 2810. It is worth noting that this figure is exemplary; other types of structures may also be used to supplement or replace this structure to implement telecommunications functions or other functions.
  • the processor 2810 may be configured to execute a program to implement the data acquisition method as described in the embodiment of the first aspect.
  • the processor 2810 may be configured to perform the following control: acquire the first data X input to the information generation unit; and acquire the second data Y corresponding to the first data X and output from the information generation unit.
  • the processor 2810 may be configured to execute a program to implement the data acquisition method described in the embodiment of the second aspect.
  • the processor 2810 may be configured to perform the following control: acquire the first data X output by the information reconstruction part; and acquire the second data X input to the information reconstruction part and corresponding to the first data X. DataY.
  • the processor 2810 may be configured to execute a program to implement the data acquisition method described in the embodiment of the fourth aspect.
  • the processor 2810 may be configured to perform the following control: obtain a codebook vector output by the CSI reconstruction unit; and obtain CSI bit information input to the CSI reconstruction unit and corresponding to the codebook vector.
  • the terminal device 2800 may also include: a communication module 2830, an input unit 2840, a display 2850, and a power supply 2860.
  • the functions of the above components are similar to those in the prior art and will not be described again here. It is worth noting that the terminal device 2800 does not have to include all the components shown in Figure 28, and the above components are not required; in addition, the terminal device 2800 can also include components not shown in Figure 28, please refer to the current There is technology.
  • An embodiment of the present application also provides a computer program, wherein when the program is executed in a terminal device, the program causes the terminal device to execute the data acquisition method described in the embodiments of the first, second, and fourth aspects.
  • Embodiments of the present application also provide a storage medium storing a computer program, wherein the computer program causes the terminal device to execute the data acquisition method described in the embodiments of the first, second, and fourth aspects.
  • An embodiment of the present application also provides a computer program, wherein when the program is executed in a network device, the program causes the network device to execute the data acquisition method described in the embodiments of the first, second, and third aspects.
  • An embodiment of the present application also provides a storage medium storing a computer program, wherein the computer program causes the network device to execute the data acquisition method described in the embodiments of the first, second, and third aspects.
  • the above devices and methods of this application can be implemented by hardware, or can be implemented by hardware combined with software.
  • the present application relates to a computer-readable program that, when executed by a logic component, enables the logic component to implement the apparatus or component described above, or enables the logic component to implement the various methods described above or steps.
  • This application also involves storage media used to store the above programs, such as hard disks, magnetic disks, optical disks, DVDs, flash memories, etc.
  • the methods/devices described in connection with the embodiments of the present application may be directly embodied as hardware, a software module executed by a processor, or a combination of both.
  • one or more of the functional block diagrams and/or one or more combinations of the functional block diagrams shown in the figure may correspond to each software module of the computer program flow, or may correspond to each hardware module.
  • These software modules can respectively correspond to the various steps shown in the figure.
  • These hardware modules can be implemented by solidifying these software modules using a field programmable gate array (FPGA), for example.
  • FPGA field programmable gate array
  • the software module may be located in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
  • a storage medium may be coupled to the processor such that the processor can read information from the storage medium and write information to the storage medium; or the storage medium may be an integral part of the processor.
  • the processor and storage media may be located in an ASIC.
  • the software module can be stored in the memory of the mobile terminal or in a memory card that can be inserted into the mobile terminal.
  • the software module can be stored in the MEGA-SIM card or the large-capacity flash memory device.
  • One or more of the functional blocks and/or one or more combinations of the functional blocks described in the accompanying drawings may be implemented as a general-purpose processor or a digital signal processor (DSP) for performing the functions described in this application. ), application specific integrated circuit (ASIC), field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, or any appropriate combination thereof.
  • DSP digital signal processor
  • ASIC application specific integrated circuit
  • FPGA field programmable gate array
  • One or more of the functional blocks and/or one or more combinations of the functional blocks described in the accompanying drawings can also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, or multiple microprocessors. processor, one or more microprocessors combined with DSP communications, or any other such configuration.
  • a data acquisition method wherein the AI/ML model includes an information generation part located in the first device and an information reconstruction part located in the second device, the method includes:
  • the second device acquires the first data X input to the information generation unit
  • the second device acquires second data Y corresponding to the first data X and output from the information generation section.
  • the second device inputs the second data Y into the information reconstruction part, and uses the first data X as label data to train the information reconstruction part.
  • the output is the second data Y or data approximate to the second data Y;
  • the trained information reconstruction unit in the second device When the second data Y is input, the trained information reconstruction unit in the second device outputs the first data X or data that is similar to the first data X.
  • model ID and/or version ID are uploaded to a core network, an external server, or a network device.
  • the second device obtains the first data X and/or the second data Y from inside the second device or from outside the second device according to the identification information related to the information generation unit.
  • identification information related to the information generation part includes: the model identification and/or version information corresponding to the information generation part, and/or the information generation part corresponds to The data configuration information of the model.
  • the second device sends an AI/ML related capability query to the first device
  • the second device receives the AI/ML related capability response fed back by the first device.
  • AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, and data configuration information. , AI/ML support training capability information, AI/ML upgrade capability information.
  • the The second device acquires the first data and/or the second data corresponding to the model identification and/or version information of the information generation unit of the first device.
  • the second device sets the model identification and/or version information of the information reconstruction unit to be the same as the information generated. have the same model ID and/or version information.
  • the second device sends confirmation information to the first device confirming that its information generation unit can be used, and/or sends instruction information to enable the information generation unit and/or the information reconstruction unit.
  • a data acquisition method wherein the AI/ML model includes an information generation part located in the first device and an information reconstruction part located in the second device, the method includes:
  • the first device acquires the first data X output by the information reconstruction unit
  • the first device acquires the second data Y input to the information reconstruction section and corresponding to the first data X.
  • the first device inputs the first data X into the information generation part, and uses the second data Y as label data to train the information generation part.
  • the output is the second data Y or data approximate to the second data Y;
  • the trained information reconstruction unit in the second device When the second data Y is input, the trained information reconstruction unit in the second device outputs the first data X or data similar to the first data X.
  • model ID and/or version ID are uploaded to a core network, an external server, or a network device.
  • the first device obtains the second data and/or the first data from inside the first device or from outside the first device according to the identification information related to the information reconstruction part.
  • identification information related to the information reconstruction part includes: the model identification and/or version information corresponding to the information reconstruction part, and/or the information reconstruction part The data configuration information of the model corresponding to the structure.
  • the first device sends an AI/ML related capability query to the second device;
  • the first device receives the AI/ML related capability response fed back by the second device.
  • AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, and data configuration information. , AI/ML support training capability information, AI/ML upgrade capability information.
  • the The first device acquires the first data and/or the second data corresponding to the model identification and/or version information of the information reconstruction unit of the second device.
  • the first device sets the model identification and/or version information of the information generation unit to be consistent with the information reconstruction unit. have the same model ID and/or version information.
  • the first device sends confirmation information to the second device confirming that its information reconstruction unit can be used, and/or instruction information to enable the information generation unit and/or the information reconstruction unit.
  • a data acquisition method wherein the AI/ML model includes a CSI generation unit located in the terminal device and a CSI reconstruction unit located in the network device, the method includes:
  • the network device acquires a specific codebook vector input to the CSI generation unit.
  • the network device acquires CSI bit information output from the CSI generation unit corresponding to the codebook vector.
  • the network device inputs the CSI bit information into the CSI reconstruction part, and uses the codebook vector as label data to train the CSI reconstruction part.
  • the network device obtains the CSI bit information and/or the codebook vector from inside the network device or from outside the network device according to the identification information related to the CSI generation unit.
  • the identification information related to the CSI generation unit includes: the model identification and/or version information corresponding to the CSI generation unit, and/or the CSI generation unit corresponding The data configuration information of the model, and/or the data index of the CSI generating unit, and/or the indication sequence of the codebook vector or DFT vector of the CSI generating unit.
  • the network device sends model training instructions and configuration information for the model identification to the terminal device;
  • the configuration information includes an instruction sequence of a training codebook vector or a DFT matrix vector;
  • the network device receives an indication sequence sent by the terminal device for indicating a codebook vector or a DFT vector and a CSI generation unit output bit sequence corresponding to the vector.
  • the network device Based on the instruction sequence, the network device generates a corresponding codebook vector or DFT matrix vector according to predefined rules.
  • the network device receives a resource request from the terminal device.
  • the network device configures uplink resources for the terminal device, so that the terminal device sends the indication sequence and the CSI generation unit outputs a bit sequence.
  • the AI/ML model is a bilateral model with a model identifier and a version identifier; the CSI generation part of the one bilateral model and the CSI reconstruction The CSI generation unit and the CSI reconstruction unit have different sub-identities.
  • the network device sends an AI/ML related capability query to the terminal device;
  • the network device receives the AI/ML related capability response fed back by the terminal device.
  • AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, and data configuration information. , AI/ML training capability information, AI/ML upgrade capability information.
  • the terminal device sends a training CSI bit sequence and/or a corresponding codebook vector corresponding to the model identification and/or version information of the CSI generation unit.
  • the network device After the CSI reconstruction unit completes training using the CSI bit information and the codebook vector, the network device sets the model identification and/or version information of the CSI reconstruction unit to be consistent with the CSI generation unit. have the same model ID and/or version information.
  • the network device sends instruction information to enable the CSI generation unit and/or the CSI reconstruction unit to the terminal device, and/or sends confirmation information confirming that the CSI generation unit can be used.
  • the network device expands the codebook vector and uses the expanded codebook vector as the label data.
  • a data acquisition method wherein the AI/ML model includes a CSI generation unit located in the terminal device and a CSI reconstruction unit located in the network device, the method includes:
  • the terminal device obtains the codebook vector output by the CSI reconstruction unit.
  • the terminal device acquires CSI bit information input to the CSI reconstruction unit and corresponding to the codebook vector.
  • the terminal device inputs the codebook vector into the CSI generation unit, and uses the CSI bit information as label data to train the CSI generation unit.
  • the CSI bit information is carried by a control channel or a data channel and transmitted via the air interface, and/or the codebook vector passes the data according to predefined rules Index or indicator sequence or Precoding Matrix Indicator (PMI) generation.
  • PMI Precoding Matrix Indicator
  • the terminal device obtains the CSI bit information and/or the codebook vector from a memory or from inside the terminal device or from outside the terminal device according to the identification information related to the CSI reconstruction unit.
  • the identification information related to the CSI reconstruction part includes: the model identification and/or version information corresponding to the CSI reconstruction part, and/or the reconstruction The data configuration information of the model corresponding to the generation unit, and/or the data index of the CSI reconstruction unit, and/or the indication sequence of the codebook vector or DFT vector of the CSI reconstruction unit.
  • the terminal device receives the model training instructions and configuration information for the model identification sent by the network device; the configuration information includes an instruction sequence of a training codebook vector or a DFT matrix vector and the CSI reconstruction unit input corresponding to the vector bit sequence.
  • the terminal device Based on the instruction sequence, the terminal device generates a corresponding codebook vector or DFT matrix vector according to predefined rules.
  • the terminal device receives the resource configuration of the network device.
  • the terminal device receives the indication sequence and the CSI reconstruction unit input bit sequence according to the resource configuration.
  • the AI/ML model is a bilateral model with a model identification and a version identification; the CSI generation part and the CSI reconstruction part of the same AI/ML model have For the same model identifier and version identifier, the CSI generation unit and the CSI reconstruction unit have different sub-identities.
  • the terminal device receives the AI/ML related capability query sent by the network device.
  • the terminal device feeds back an AI/ML related capability response to the network device.
  • AI/ML related capabilities include at least one of the following: signal processing module information, AI/ML support information, AI/ML model identification information, version information, and data configuration information , AI/ML training capability information, AI/ML upgrade capability information.
  • the terminal device After the CSI generation unit completes training using the codebook vector and the CSI bit information, the terminal device sets the model identification and/or version information of the CSI generation unit to be the same as those of the CSI reconstruction unit.
  • the model ID and/or version information are the same.
  • the terminal device sends confirmation information to the network device confirming that its CSI generation unit can be used, and/or indication information indicating that the CSI generation unit is updated.
  • the terminal device expands the codebook vector and uses the expanded codebook vector as the label data.
  • a network device comprising a memory and a processor, the memory stores a computer program, and the processor is configured to execute the computer program to implement the data acquisition method as described in any one of appendices 1 to 56 .
  • a terminal device comprising a memory and a processor
  • the memory stores a computer program
  • the processor is configured to execute the computer program to implement any one of appendices 1 to 38, 57-74 data acquisition method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Power Engineering (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请实施例提供一种数据获取方法和装置。AI/ML模型包括位于第一设备内的信息生成部和位于所述第二设备内的信息重构部，所述方法包括：第二设备获取被输入所述信息生成部的第一数据X；以及获取与所述第一数据X对应并从所述信息生成部输出的第二数据Y。

Description

数据获取方法与装置 技术领域
本申请实施例涉及通信技术领域。
背景技术
多天线技术在LTE、LTE-A和5G NR系统中得到了广泛应用。特别是大规模天线技术(masssive MIMO)在5G标准中得到了非常好的支持。可以预见在5G-Advanced阶段和6G阶段，massive MIMO技术将得到更为广泛的利用，增强性能的大规模及超大规模天线技术也是下一代移动通信系统的研究重点。
伴随着人工智能(AI，Artificial Intelligence)和机器学习(ML，Machine Learning)技术的发展，将AI/ML技术应用到无线通信上，来解决传统方法的难点成为当前一个技术方向。AI/ML模型应用于无线通信系统，特别是应用于空口的传输是5G-Advanced以及6G阶段的新技术。
例如，对于信道状态信息(CSI，Channel State Information)的测量和反馈，在终端设备侧测量CSI，并使用AI/ML模型生成CSI反馈信息，通过空口发送给网络侧后，网络侧接收该CSI反馈信息，并通过相应的AI/ML模型恢复出原始的CSI。这样的例子中，通过AI/ML模型可以降低CSI反馈开销，或者提升反馈的质量，进而达到通信质量的提升。
类似地，发送端(网络设备或终端设备)的信息经过AI/ML模型处理后，通过空口发送给接收端(终端设备或网络设备)，接收端使用与发送端AI/ML模型相对应的模型恢复出发送端输入AI/ML模型的信息，通常可以将这类AI/ML模型称为双边模型(two-sided model)。
应该注意，上面对技术背景的介绍只是为了方便对本申请的技术方案进行清楚、完整的说明，并方便本领域技术人员的理解而阐述的。不能仅仅因为这些方案在本申请的背景技术部分进行了阐述而认为上述技术方案为本领域技术人员所公知。
发明内容
但是，发明人发现：双边模型的成对属性使得其训练的过程，需要空口两侧的网络设备和终端设备进行协作，经过空口进行网络模型训练，这会造成巨大的空口开销。
针对上述问题的至少之一，本申请实施例提供一种数据获取方法，其中，AI/ML模型包括位于第一设备内的信息生成部和位于第二设备内的信息重构部，所述方法包括：
所述第二设备获取被输入所述信息生成部的第一数据X；以及
所述第二设备获取与所述第一数据X对应并从所述信息生成部输出的第二数据Y。
根据本申请实施例的一个方面，提供一种数据获取装置，配置于第二设备；其中，AI/ML模型包括位于第一设备内的信息生成部和位于所述第二设备内的信息重构部，所述数据获取装置包括：
第一获取单元，其获取被输入所述信息生成部的第一数据X；以及
第二获取单元，其获取与所述第一数据X对应并从所述信息生成部输出的第二数据Y。
根据本申请实施例的另一个方面，提供一种数据获取方法，其中，AI/ML模型包括位于第一设备内的信息生成部和位于第二设备内的信息重构部，所述方法包括：
所述第一设备获取所述信息重构部输出的第一数据X；以及
所述第一设备获取被输入到所述信息重构部的、并且与所述第一数据X对应的第二数据Y。
根据本申请实施例的另一个方面，提供一种数据获取装置，配置于第一设备；其中，AI/ML模型包括位于所述第一设备内的信息生成部和位于第二设备内的信息重构部，所述装置包括：
第三获取单元，其获取所述信息重构部输出的第一数据X；以及
第四获取单元，其获取被输入到所述信息重构部的、并且与所述第一数据X对应的第二数据Y。
根据本申请实施例的另一个方面，提供一种通信系统，其中，AI/ML模型包括位于终端设备内的CSI生成部和位于网络设备内的CSI重构部，所述系统包括：
网络设备，其获取被输入所述CSI生成部的特定的码书向量；以及获取与所述码书向量对应的从所述CSI生成部输出的CSI比特信息；
终端设备，其获取所述CSI重构部输出的码书向量；以及获取被输入所述CSI重构部的、并且与所述码书向量对应的CSI比特信息。
本申请实施例的有益效果之一在于：第二设备获取被输入信息生成部的第一数据X；以及第二设备获取与所述第一数据X对应并从所述信息生成部输出的第二数据Y。由此，能够以较小的代价获得合适的数据，从而支持双边网络模型架构。
参照后文的说明和附图，详细公开了本申请的特定实施方式，指明了本申请的原理可以被采用的方式。应该理解，本申请的实施方式在范围上并不因而受到限制。在所附权利要求的精神和条款的范围内，本申请的实施方式包括许多改变、修改和等同。
针对一种实施方式描述和/或示出的特征可以以相同或类似的方式在一个或更多个其它实施方式中使用，与其它实施方式中的特征相组合，或替代其它实施方式中的特征。
应该强调，术语“包括/包含”在本文使用时指特征、整件、步骤或组件的存在，但并不排除一个或更多个其它特征、整件、步骤或组件的存在或附加。
附图说明
在本申请实施例的一个附图或一种实施方式中描述的元素和特征可以与一个或更多个其它附图或实施方式中示出的元素和特征相结合。此外，在附图中，类似的标号表示几个附图中对应的部件，并可用于指示多于一种实施方式中使用的对应部件。
图1是本申请实施例的通信系统的示意图；
图2是本申请实施例的双边模型的一个示例图；
图3是本申请实施例的双边模型的另一个示例图；
图4是本申请实施例的双边模型的另一个示例图；
图5是本申请实施例的双边模型的另一个示例图；
图6是本申请实施例的数据获取方法的一示意图；
图7是本申请实施例的数据的一示意图；
图8是本申请实施例的联合训练的一示意图；
图9是本申请实施例的分离训练的一示意图；
图10是本申请实施例的经过分离训练后的一示意图；
图11是本申请实施例的一个双边模型在gNB和UE应用的示意图
图12是本申请实施例的数据获取方法的一示意图；
图13是本申请实施例的构建数据集的一示例图；
图14是本申请实施例的数据获取方法的一示意图；
图15是本申请实施例的CSI重构部获取数据的一示例图；
图16是本申请实施例的CSI重构部进行训练的一示例图；
图17是本申请实施例的CSI重构部获取数据的另一示例图；
图18是本申请实施例的CSI重构部进行训练的另一示例图；
图19是本申请实施例的CSI重构部进行训练的另一示例图；
图20是本申请实施例的CSI重构部训练的一信令示意图；
图21是本申请实施例的数据获取方法的一示意图；
图22是本申请实施例的CSI生成部获取数据的一示例图；
图23是本申请实施例的CSI生成部进行训练的另一示例图；
图24是本申请实施例的CSI生成部训练的一信令示意图；
图25是本申请实施例的数据获取装置的一示意图；
图26是本申请实施例的数据获取装置的另一示意图；
图27是本申请实施例的网络设备的一示意图；
图28是本申请实施例的终端设备的一示意图。
具体实施方式
参照附图，通过下面的说明书，本申请的前述以及其它特征将变得明显。在说明书和附图中，具体公开了本申请的特定实施方式，其表明了其中可以采用本申请的原则的部分实施方式，应了解的是，本申请不限于所描述的实施方式，相反，本申请包括落入所附权利要求的范围内的全部修改、变型以及等同物。
在本申请实施例中，术语“第一”、“第二”等用于对不同元素从称谓上进行区分，但并不表示这些元素的空间排列或时间顺序等，这些元素不应被这些术语所限制。术语“和/或”包括相关联列出的术语的一种或多个中的任何一个和所有组合。术语“包含”、“包括”、“具有”等是指所陈述的特征、元素、元件或组件的存在，但并不排除存在或添加一个或多个其他特征、元素、元件或组件。
在本申请实施例中，单数形式“一”、“该”等包括复数形式，应广义地理解为“一种”或“一类”而并不是限定为“一个”的含义；此外术语“所述”应理解为既包括单数形式也包括复数形式，除非上下文另外明确指出。此外术语“根据”应理解为“至少部分根 据……”，术语“基于”应理解为“至少部分基于……”，除非上下文另外明确指出。
在本申请实施例中，术语“通信网络”或“无线通信网络”可以指符合如下任意通信标准的网络，例如长期演进(LTE，Long Term Evolution)、增强的长期演进(LTE-A，LTE-Advanced)、宽带码分多址接入(WCDMA，Wideband Code Division Multiple Access)、高速报文接入(HSPA，High-Speed Packet Access)等等。
并且，通信系统中设备之间的通信可以根据任意阶段的通信协议进行，例如可以包括但不限于如下通信协议：1G(generation)、2G、2.5G、2.75G、3G、4G、4.5G以及5G、新无线(NR，New Radio)、未来的6G等等，和/或其他目前已知或未来将被开发的通信协议。
在本申请实施例中，术语“网络设备”例如是指通信系统中将终端设备接入通信网络并为该终端设备提供服务的设备。网络设备可以包括但不限于如下设备：基站(BS，Base Station)、接入点(AP、Access Point)、发送接收点(TRP，Transmission Reception Point)、广播发射机、移动管理实体(MME、Mobile Management Entity)、网关、服务器、无线网络控制器(RNC，Radio Network Controller)、基站控制器(BSC，Base Station Controller)等等。
其中，基站可以包括但不限于：节点B(NodeB或NB)、演进节点B(eNodeB或eNB)以及5G基站(gNB)，IAB宿主等等，此外还可包括远端无线头(RRH，Remote Radio Head)、远端无线单元(RRU，Remote Radio Unit)、中继(relay)或者低功率节点(例如femeto、pico等等)。并且术语“基站”可以包括它们的一些或所有功能，每个基站可以对特定的地理区域提供通信覆盖。术语“小区”可以指的是基站和/或其覆盖区域，这取决于使用该术语的上下文。
在本申请实施例中，术语“用户设备”(UE，User Equipment)或者“终端设备”(TE，Terminal Equipment或Terminal Device)例如是指通过网络设备接入通信网络并接收网络服务的设备。终端设备可以是固定的或移动的，并且也可以称为移动台(MS，Mobile Station)、终端、用户台(SS，Subscriber Station)、接入终端(AT，Access Terminal)、站，等等。
其中，终端设备可以包括但不限于如下设备：蜂窝电话(Cellular Phone)、个人数字助理(PDA，Personal Digital Assistant)、无线调制解调器、无线通信设备、手持设备、机器型通信设备、膝上型计算机、无绳电话、智能手机、智能手表、数字相机， 等等。
再例如，在物联网(IoT，Internet of Things)等场景下，终端设备还可以是进行监控或测量的机器或装置，例如可以包括但不限于：机器类通信(MTC，Machine Type Communication)终端、车载通信终端、设备到设备(D2D，Device to Device)终端、机器到机器(M2M，Machine to Machine)终端，等等。
此外，术语“网络侧”或“网络设备侧”是指网络的一侧，可以是某一基站，也可以包括如上的一个或多个网络设备。术语“用户侧”或“终端侧”或“终端设备侧”是指用户或终端的一侧，可以是某一UE，也可以包括如上的一个或多个终端设备。本文在没有特别指出的情况下，“设备”可以指网络设备，也可以指终端设备。
以下通过示例对本申请实施例的场景进行说明，但本申请不限于此。
图1是本申请实施例的通信系统的示意图，示意性说明了以终端设备和网络设备为例的情况，如图1所示，通信系统100可以包括网络设备101和终端设备102、103。为简单起见，图1仅以两个终端设备和一个网络设备为例进行说明，但本申请实施例不限于此。
在本申请实施例中，网络设备101和终端设备102、103之间可以进行现有的业务或者未来可实施的业务发送。例如，这些业务可以包括但不限于：增强的移动宽带(eMBB，enhanced Mobile Broadband)、大规模机器类型通信(mMTC，massive Machine Type Communication)和高可靠低时延通信(URLLC，Ultra-Reliable and Low-Latency Communication)，等等。
值得注意的是，图1示出了两个终端设备102、103均处于网络设备101的覆盖范围内，但本申请不限于此。两个终端设备102、103可以均不在网络设备101的覆盖范围内，或者一个终端设备102在网络设备101的覆盖范围之内而另一个终端设备103在网络设备101的覆盖范围之外。
在本申请实施例中，高层信令例如可以是无线资源控制(RRC)信令；例如称为RRC消息(RRC message)，例如包括MIB、系统信息(system information)、专用RRC消息；或者称为RRC IE(RRC information element)。高层信令例如还可以是MAC(Medium Access Control)信令；或者称为MAC CE(MAC control element)。但本申请不限于此。
在本申请实施例中，网络设备和/或终端设备中可以配置并运行一个或多个 AI/ML模型。AI/ML模型可以用于无线通信的各种信号处理功能，例如CSI估计和上报、波束管理和波束预测等等；本申请不限于此。
图2是本申请实施例的双边模型的一个示例图。如图2所示，AI/ML模型的一部分在发送端TX，另一部分在接收端RX。在发送端，源信息X输入模型后生成信息Y，信息Y被传输到接收端。在接收端，经由信道传输后的信息Y’被输入模型后生成恢复信息X’。以上信息满足如下公式1：
Figure PCTCN2022112299-appb-000001
其中，f(·)是TX侧神经网络的传递函数，X是TX处神经网络的源信息，Y是其输出信号，h(·)是TX和RX之间的通道响应系数，n是高斯噪声，g(·)是RX侧神经网络的传递函数，Y'是神经网络在RX处的输入信号，X'是其输出信号。
在迭代训练过程中，损失函数使X'尽可能接近X，即‖X-X'‖→0。损失函数可以使用余弦相似度、NMSE或MSE作为度量。训练完成后，TX神经网络和RX神经网络分别部署在TX侧(例如UE侧)和RX侧(例如gNB侧)，通过空口(air interface)相互通信。
例如，双边模型可以应用于CSI生成和上报。
图3是本申请实施例的双边模型的另一个示例图。如图3所示，双边模型由终端设备侧的CSI生成部(也可称为CSI encoder)和网络设备侧的CSI重构部(也可称为CSI decoder)组成。
如图3所示，CSI生成部的输入是信道系数信息，可以是由信道估计产生的信道系数向量或矩阵，或者是由上述信道系数矩阵经过矩阵分解产生的信道特征向量或奇异向量。CSI生成部的输出是经过特征提取、信息压缩，并经过量化过程而形成的CSI比特序列。该比特序列作为上行控制信息(UCI)的一部分，承载在上行控制信道或者上行数据信道上，发送给网络侧。
如图3所示，网络设备从接收的UCI中提取出相应的CSI比特序列，输入到CSI重构部。该CSI重构部的输出为重构的信道系数向量或矩阵，或者信道特征向量或奇异向量，具体是哪一种与CSI生成部的输入相对应。
除了上述的CSI例子，双边模型还可替代传统信号处理方法，用于通信系统中各个成对处理的模块当中。
图4是本申请实施例的双边模型的另一个示例图。如图4所示，例如，信源信号的源编码和解码(source coding/decoding)，信道编码和解码(channel coding/decoding)，信道调制和解调(modulation/demodulation)，多天线模块mimo/de-mimo(具体可包含针对mimo多数据流复用，mimo空时编码，mimo precoding/beamforming等相应的发收两端的成对信号处理，其他发送端由多个数据流到多个天线端口的发送网络部，接收端的对应的接收网络部)，发送波形的成形和解波形成形(waveforming/de-waveforming)，等等。
因此，可以将通信两端的发送端和接收端的信号处理流程或功能都合并，并引入神经网络模型来实现原有的信号处理。
图5是本申请实施例的双边模型的另一个示例图。收发两端可如图5所示，在发送端设置信息生成部，在接收端设置信息重构部，从而可利用双边模型来进行成对的信号处理。
理想情况下，这种双边模型的构建、升级，需要通过联合训练的方式来完成，即需要信息生成部和信息重构部一起完成联合初始训练，联合fine tuning，联合再训练，联合的模型升级。上述的双边模型所有联合训练相关过程包含离线操作或者在线操作。
而实际中，无论是离线的联合训练还是在线的联合训练都面临巨大挑战。因为通常双边模型的信息生成部和信息重构部由不同的厂商来开发，而且是部署在空口两端的。离线训练面临厂商间协作问题，在线训练面临空口开销问题，另外还涉及硬件支持问题，模型的知识产权等问题。
在双边模型构建阶段，网络厂商的设备很难和终端产商设备一起工作来完成对一个双边模型的训练、验证和测试的过程。如果两者分别进行模型开发，则分别开发完成的模型如何实现协作工作是一个难题。
另外，发明人发现，对于双边模型，如果网络侧的模型和终端侧的模型不匹配，则通信两侧不能联合工作。在一种方案中，一侧可以传输其匹配的模型给另一侧。比如，网络侧把与CSI重构部匹配的CSI生成部通过空口传输给终端侧，终端侧接收后使用该CSI生成部来工作。但是，这要求终端侧具有AI/ML接收及升级的能力，对其硬件软件的要求非常高。另一方面，考虑AI/ML模型的知识产权归属问题，也使 得AI/ML模型在网络设备和终端设备传输使用的方案，存在设备能力之外的知识产权问题及其他厂商之间的协作问题。因此，通过模型传输来达到双边模型配对的方案会面临非常多的问题。
在另一种方案中，可以更新终端侧或者网络侧的AI/ML模型，使两侧模型能够配对并联合工作。为实现这一过程，需要对一侧或者两侧的AI/ML模型进行联合训练。模型训练需要数据集，对于监督学习还需要标签数据集。对于部署在网络侧和终端侧的AI/ML模型，联合训练带来非常大的空口开销，包括输出相关数据集的开销，来回迭代过程对应的输入输出的开销等，很难在移动通信的空口支持这样的过程。
因此，为了支持双边网络模型架构，需要有一种机制解耦联合训练，双边模型的每一边可以独立训练。为此本申请实施例提出使用共用成对数据集来实现独立训练的方案，并给出减少共用数据集尺度及减少获取数据空口开销的方法和装置。
对于本申请实施例中所涉及的术语有以下说明：
联合训练(joint training)或者称为联合学习，包括双边模型的两侧模型部需要联合进行训练的过程(包含初始训练、再训练、在线训练、离线训练、fine tuning等，也包含训练模型中必要的模型验证、模型测试的子过程)。
分离训练或者称为分别训练或者称为独立训练，包括双边模型的两侧模型部可以单独进行训练的过程(包含初始训练、再训练、在线训练、离线训练、fine tuning等，也包含训练模型中必要的模型验证、模型测试的子过程)。
以上示意性说明了相关概念和术语，但本申请不限于此，也可替换为其他概念和术语。关于涉及的具体内容还可以参考相关技术。
第一方面的实施例
本申请实施例提供一种数据获取方法，从第二设备侧进行说明。AI/ML模型包括位于第一设备内的信息生成部和位于第二设备内的信息重构部。该第二设备可以是网络设备，第一设备为终端设备；该第二设备也可以是终端设备，第一设备为网络设备。
图6是本申请实施例的数据获取方法的一示意图，如图6所示，该方法包括：
601，第二设备获取被输入信息生成部的第一数据X；以及
602，第二设备获取与第一数据X对应并从所述信息生成部输出的第二数据Y。
值得注意的是，以上附图6仅对本申请实施例进行了示意性说明，但本申请不限 于此。例如可以适当地调整各个操作之间的执行顺序，此外还可以增加其他的一些操作或者减少其中的某些操作。本领域的技术人员可以根据上述内容进行适当地变型，而不仅限于上述附图6的记载。
在一些实施例中，针对不同信号处理功能可以分别运行AI/ML模型。例如，针对CSI上报的AI/ML模型，可以具有不同的模型组标识、模型标识和版本标识。针对波束管理的AI/ML模型，可以具有另外的模型组标识、模型标识和版本标识。
在一些实施例中，所述AI/ML模型为双边模型，具有模型标识和版本标识；同一AI/ML模型的信息生成部和信息重构部使用相同的模型标识和版本标识，所述信息生成部和所述信息重构部具有不同的子标识。
例如，一个双边模型具有一个模型标识(可以是一级标识或多级标识)和版本标识，可以区分两个不同的双边模型(例如存在层数、节点数、超参数、输入输出信号格式、预处理及后处理配置等的不同)。同一个双边模型的信息生成部和信息重构部，具有同样的模型标识和版本标识。此外，可以用子标识区分是信息生成部还是信息重构部。
例如，模型标识为一级标识XXXXX(单级model标识).Vxx(版本标识)；
再例如，模型标识为多级标识：第一级XXX-第二级XXXX-第n级XXXX(多级model标识).Vxxx(版本标识)。
再例如，在模型标识(含版本标识)之外，可以用1比特子标识区分生成部和重构部。例如，信息生成部为：模型标识.版本标识.0；信息重构部为：模型标识.版本标识.1。
图7是本申请实施例的数据的一示意图，如图7所示，对于一个双边模型，虽然被分开部署在发送侧和接收侧，但信息生成部的输入和输出对应信息重构部的输出和输入。理想的，
X＝X′；Y＝Y′
表1示出了本申请实施例的数据的示例。
表1
  信息生成部输入 信息生成部输出 信息重构部输入 信息重构部输出
模型标识 {X(数据索引|数据)} {Y(数据索引|数据)} {Y(数据索引|数据)} {X(数据索引|数据)}
如表1所示，对应某一模型标识(含版本标识)的双边模型，其信息生成部和信息重构部具有相同(或近似)对偶的训练数据集。即：可以把信息生成部输入数据集{X}及对应的输出数据集{Y}共享给信息重构部，信息重构部输入{Y}且输出{X}。因此，这种特征可以用于分离训练，
在一些实施例中，所述第二设备将所述第二数据Y输入到所述信息重构部中，并将所述第一数据X作为标签数据对所述信息重构部进行训练。
由此，能够以较小的代价获得合适的数据，从而支持双边网络模型架构。
在一些实施例中，所述第一设备内的所述信息生成部和所述第二设备内的所述信息重构部使用所述第一数据X和所述第二数据Y分别进行训练；所述第一设备内已训练的信息生成部在输入所述第一数据X时，输出为所述第二数据Y或为近似所述第二数据Y的数据；所述第二设备内已训练的信息重构部在输入所述第二数据Y时，输出为所述第一数据X或为近似所述第一数据X的数据。
例如，对于双边模型(two-sided model)，网络设备中的信息重构部和终端设备的信息生成部，分别使用同样的成对数据集(common paired datasets){X}和{Y}进行训练。在训练后，网络设备的信息重构部在输入{Y}时，输出为{X}或近似为{X}；终端设备的信息生成部在输入{X}时，输出为{Y}或近似为{Y}；
其中X和Y的格式是预定义或配置的，如符号格式、比特数等。
在一些实施例中，所述第一数据X和所述第二数据Y是成对的数据集；所述成对的数据集具有模型标识信息。其中，一个第一数据X可以对应多个第二数据Y。
例如，一个X可以对应多个Y(如Y1，Y2)，Y1,Y2的格式是预定义的或配置的。成对数据集为：{{X}{Y1}}，{{X}{Y2}}，或{{X}{Y1}{Y2}}，例如可以在多个第一设备和一个第二设备中使用。
图8是本申请实施例的联合训练的一示意图。
在一些实施例中，如图8所示，为了使双边模型能够联合工作，信息生成部的模型需要和信息重构部的模型联合训练，经过多次的迭代，后向传播，参数升级等步骤，完成训练。
图9是本申请实施例的分离训练的一示意图。
为了实现双边模型的分离训练，预先准备了所述第一数据X和所述第二数据Y，X和Y是成对关系，多个X及对应的多个Y构成成对数据集，其中X和Y的格式 是预定义或配置的，如符号格式、比特数等。一个X可以对应多个Y(如Y1，Y2)，Y1,Y2的格式是预定义的或配置的。例如，成对数据集为：{{X}{Y1}}，{{X}{Y2}}，或{{X}{Y1}{Y2}}。如图9所示，所准备的数据集可以从各种典型信道中经过仿真分析提取出来。
例如，一个X可以对应多个Y(如Y1，Y2)，Y1,Y2的格式是预定义的或配置的，例如符号格式，比特数目，对应的port数，port配置，数据长度，维度等。成对数据集为：{{X}{Y1}}，{{X}{Y2}}，或{{X}{Y1}{Y2}}，例如可以在多个第一设备和一个第二设备中使用。
在如图9所示的示例中，其中共同成对数据集的{X}作为UE侧模型部的输入数据，{Y}作为输出的参考数据或标签数据，训练的目的是使得训练后，输入每一个X时输出是Y，或者非常接近Y。同样的数据集也共享给gNB侧模型部，{Y}作为gNB侧模型部的输入数据，{X}作为输出的参考数据或标签数据，训练的目的是使得训练后，输入每一个Y时输出是X，或者非常接近X。
因为采用了上述的成对数据集分别对UE侧模型部(比如称为信息生成部)和gNB侧模型部(比如称为信息重构部)进行训练，基于上述训练要求，单独训练完成后，当联合在一起工作时，至少对输入数据集{X}，可以实现与联合训练一样的数据恢复效果。
图10是本申请实施例的经过分离训练后的一示意图。如图10所示。当数据集{X}足够表征模型所应用无线信道的特征时，在其应用到通信系统时，就能够非常好的工作。
另外，可以通过更新该共用的成对数据集，在模型部署到网络设备和终端设备后，在运行过程中，利用基于信道更新的数据集，分别进行模型再训练或者fine tuning，来保证模型对新应用环境的适应性。经过不断地更新成对数据集，并以此来分别对成对的双边模型训练的方法，可以使得双边模型联合训练的困难得以解决。
上述方法可以用于模型的初始训练，再训练，离线训练，在线训练，fine tuning等训练过程。另外一种情况中，如果不能在各厂商间建立基于共用数据集的生态，可以采用基于online、基于空口的方式来进行。
图11是本申请实施例的一个双边模型在gNB和UE应用的示意图，其中考虑了很多可能的情况。例如，模型监控发现性能不好时，可以启动模型的训练过程，两侧 模型部通过新收集的数据，基于上述共用成对数据集的原则进行交互和独立训练，训练后的模型可以通过空口的使用和性能监测，在性能达标后可作为更新的模型来进行model inference。这样避免了联合训练和模型传输面临的问题。
经过训练并空口测试的模型可以注册信道模型ID或新版本ID，可以上传到核心网，远端服务器，OTT服务器，或传递给其他gNB，等等。
以上示意性说明了共享数据以及分离训练等，本申请不限于此。以下再对如何获取数据进行示意性说明。
在一些实施例中，所述第一数据X为输入到所述信息生成部的特定数据集的部分或全部数据；所述第二数据Y为将所述第一数据X输入到所述信息生成部后生成的对应数据。
例如，特定数据集可以是对应于当前生成部模型的、基于预定义规则可以生成的数据集，或者可以是所有信息生成部模型共用的数据集，例如所有可能的信道信息量化向量。
在一些实施例中，所述第二设备根据与所述信息生成部相关的标识信息从所述第二设备内部或从所述第二设备外部获得所述第一数据X和/或所述第二数据Y。
例如，第二设备内部可以包括内存、磁盘等存储器件；第二设备外部可以包括与第二设备相连接的设备，例如核心网、云端、第三方服务器、OTT服务器、云端存储设备等等。
在一些实施例中，所述与信息生成部相关的标识信息包括：所述信息生成部对应的模型标识和/或版本信息，和/或，所述信息生成部对应的模型的数据配置信息。本申请不限于此，例如还可以是其他的标识。
在一些实施例中，所述第二数据Y通过控制信道或数据信道承载并经由空口传输，或者，所述第二数据Y根据预定义规则通过数据索引生成，或者第二数据Y的索引由所述第一设备通过空口发送给所述第二设备。
在一些实施例中，所述第一数据X被预先存储在所述第二设备内，或者，所述第一数据X根据预定义规则通过数据索引生成，或者，所述第一数据X由所述第一设备通过空口发送给所述第二设备，或者第一数据X的索引由所述第一设备通过空口发送给所述第二设备。
以下通过一些例子进行说明。假定信息生成部或信息重构部进行初始训练或离线 训练。信息生成部和信息重构部约定好输入数据配置和对应的输出数据配置。配置可以包括数据的维度，格式等。然后分别进行训练，具体训练方法不限。
例如，信息生成部先进行训练，为了方便训练，其过程中可以使用与其配对的模型部来一起训练。在训练完成后，信息生成部把其输入数据集和对应的最后的输出数据集的全部或部分集合共享给重构部。信息重构部使用其输出数据集作为输入，使用其输出数据集作为输出的标签集。经过训练，使得其模型输出逼近标签数据集。这样完成了信息重构部向信息生成部的模型配对。
反之也是类似，可以先训练信息重构部，在其训练过程中可以使用一个与其配对的模型部方便训练，训练完成后，信息重构部将其输入输出数据集的全部或部分集合共享给信息生成部，用于信息生成部的训练，使得信息生成部的输出逼近信息重构部的输入。
以上过程可以通过离线的形式完成。上述训练的先、后描述只是为了叙述方便，也可以两侧同时来进行。信息生成部和信息重构部的输入输出格式经过约定或预定义，可以对齐输入输出格式，这样两者可以分别训练，训练后相互共享输入输出数据集全部集，或部分集，组成聚合的数据集，用于两侧的再次训练。
再例如，信息生成部输入数据集和信息重构部输出数据集采用共同的数据集(该共同数据集可以是双方预先有共识的，或者预先存储的，或者数据集可以通过预定义的方法生成的，等等)。这样，信息生成部利用该共同数据集作为输入数据集训练，在训练完成后，将其输出数据集共享给信息重构部，作为信息重构部的输入数据集，信息重构部将该共同数据集作为标签数据集，训练的目的是使得在该输入数据集下的输出数据集逼近标签数据集。完成信息重构部向信息生成部的配对。信息生成部向信息重构部的配对过程类似，不再赘述。
在本申请实施例中，上述信息生成部和信息重构部的数据集共享方式不做限制，可以通过远程网络传输等方式实现。在上述可能方式完成配对后，其相应的信息生成部和信息重构部就形成了一个双边模型。
可以在经历相关RAN4测试后或者其他预定义的测试，当性能满足要求后，该模型可以在网络侧注册相应的模型ID和版本ID。其所使用的训练用数据集，共同数据集可以与该模型ID绑定。由此，能够以较小的代价获得合适的数据，从而支持双边网络模型架构。所述数据集也可以上传到网络侧。通过模型ID的共享，也可达到共 享数据集的目的。
对于已经部署的AI/ML模型的网络设备和终端设备，其运行中会遇到模型不匹配的问题。比如一个终端设备在移动到某一小区后，网络侧或终端侧发现两者所对应的双边模型ID不一样，或者ID一样但版本信息不一样。可以通过空口或其他方式实现配对。
例如，可以共享X和Y。发送设备(终端设备或网络设备)可以将信息生成部的输入数据集{X(数据配置，数据索引|数据)}和输出数据{Y(数据配置，数据索引|数据)}的全部或部分，通过空口发送到接收设备(网络设备或终端设备)。该数据集可以通过控制信道或数据信道承载。
接收设备的信息重构部利用{Y(数据配置，数据索引|数据)}作为其输入数据，对其神经网络模型进行训练，并采用{X(数据配置，数据索引|数据)}作为标签(label)数据，训练的目的是使得在信息重构部输入y(数据配置，数据索引|数据)时，其中y(数据配置，数据索引|数据)是{Y(数据配置，数据索引|数据)}中的一个样本，信息重构部的输出等于或近似等于x(数据配置，数据索引|数据)，其中x(数据配置，数据索引|数据)为{X(数据配置，数据索引|数据)}中的一个样本。
通过上述方式，可以实现信息生成部不变，信息重构部的模型再训练和更新，如果模型再训练的性能满足要求，例如在经过或无需经过空口的进一步测试后，信息重构部将对应该再训练和更新的模型标记为该model ID(含版本标识)。
再例如，可以共享Y。发送设备(如终端设备或网络设备)将信息生成部针对输入为{X(数据配置，数据索引|数据)}的输出{Y(数据配置，数据索引|数据)}的全部数据集或部分数据集通过空口发送到接收设备(如网络设备或终端设备)。该数据集可以通过控制信道或数据信道承载，
接收设备的信息重构部利用{Y(数据配置，数据索引|数据)}作为其输入数据，对其神经网络模型进行训练，并采用{X(数据配置，数据索引|数据)}作为label数据，训练的目的是使得在重构部输入y(数据配置，数据索引|数据)时，其中y(数据配置，数据索引|数据)为{Y(数据配置，数据索引|数据)}中的一个样本，信息重构部的输出等于或近似等于x(数据配置，数据索引|数据)，其中x(数据配置，数据索引|数据)为{X(数据配置，数据索引|数据)}中的一个样本。
对于和{Y(数据配置，数据索引|数据)}全部集或部分集对应的{X(数据配置，数据 索引|数据)}全部集或部分集，接收设备获取的方式可以是已经存储在接收设备，通过model ID获取，或者通过model ID从远端的服务器获取，或者，根据预定义的规则由数据索引生成对应的数据。
通过上述方式，可以实现信息生成部不变，信息重构部的模型再训练和更新，如果模型再训练的性能满足要求，可以在经过空口(或无需经过)的进一步测试后，信息重构部将对应该再训练和更新的模型标记为该model ID(含版本标识)。
再例如，发送设备(如终端设备或网络设备)将信息生成部所对应的模型ID(包括版本)，和/或，模型对应输入数据配置信息通过空口发送到接收设备(如网络设备或终端设备)。模型ID和/或模型对应输入数据配置信息可以通过控制信道或数据信道承载。接收设备的信息重构部的模型ID如果和信息生成部对应的模型ID不一致，接收设备可以通过其内存获取该信息生成部模型ID和/或模型输入数据配置信息对应的{Y(数据配置，数据索引|数据)}和{X(数据配置，数据索引|数据)}，或者通过信息生成部模型ID和/或模型输入数据配置信息从远端的服务器获取。或者，信息生成部模型ID对应的{Y(数据配置，数据索引|数据)}由内存获得或从远端服务器下载，{X(数据配置，数据索引|数据)}根据预定义的规则由数据索引生成对应的数据。
通过上述方式，可以实现信息生成部不变，信息重构部的模型再训练和更新，如果模型再训练的性能满足要求，在经过空口(或无需经过)的进一步测试后，信息重构部将对应该再训练和更新的模型标记为该模型ID(含版本标识)。
以下再对本申请实施例的信令过程进行说明。
在一些实施例中，所述第二设备向所述第一设备发送AI/ML相关能力查询；以及所述第二设备接收所述第一设备反馈的AI/ML相关能力响应。其中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML支持训练能力信息、AI/ML升级能力信息。
在一些实施例中，在所述第二设备确定所述第一设备具有AI/ML能力但所述信息生成部的模型标识和/或版本信息与所述信息重构部的模型标识和/或版本信息不同时，所述第二设备获取与所述第一设备的信息生成部的模型标识和/或版本信息对应的所述第一数据和/或所述第二数据。
在一些实施例中，所述信息重构部在利用所述第一数据和所述第二数据完成训练后，所述第二设备将所述信息重构部的模型标识和/或版本信息设置为与所述信息生 成部的模型标识和/或版本信息相同。
在一些实施例中，所述第二设备向所述第一设备发送确认其信息生成部可以使用的确认信息,和/或，发送使能所述信息生成部和/或所述信息重构部的指示信息。
以信息生成部部署在终端设备，信息重构部部署在网络设备为例。
在一个示例中，对于使用双边模型的信号处理模块，网络设备向终端设备进行AI/ML相关能力问询，该问询可以包含如下的一个或多个：信号处理模块信息，AI/ML支持信息，AI/ML模型标识信息、版本信息，AI/ML模型数据配置信息，等等。终端设备进行相应的应答汇报。
例如，当网络设备发现终端设备汇报对应某模块具有AI/ML能力，但其具有的信息生成部的model ID与网络设备的信息重构部的model ID不同时，网络设备可以配置终端设备上报：在预定义的X1数据集或预先配置的第一数据集作为信息生成部输入时，对应输出的Y1数据集的数据配置信息，例如大小、结构等信息。上述信息汇报的格式是预定义的。
再例如，当网络设备发现终端设备汇报对应某模块具有AI/ML能力，其具有的信息生成部的model ID与网络设备的信息重构部的model ID相同但版本不同时。网络设备配置终端设备上报：在预定义的X11数据集或预先配置的第一数据集作为生成部输入时，对应输出的Y11数据集的数据配置信息，例如包含数据集大小、数据集结构信息(如多个数据集块)等信息。上述信息汇报的格式是预定义的。
例如，X11,Y11分别是X1，Y1的一部分。
网络设备根据终端设备上报的数据配置信息，配置或调度终端设备上报Y1数据集或Y11数据集。Y1数据集或Y11数据集由包含在上行控制信息(UCI)中，或者由AI/ML数据辅助信息上报(通过RRC消息)。
网络设备接收终端设备上报的Y1或Y11数据集后，使用Y1或Y11数据集作为其信息重构部的输入数据。网络设备存储有对应Y1或Y11数据集的X1或X11数据集，或者网络设备可以从远端服务器下载对应的X1或X11数据集。网络设备的信息重构部使用X1或X11数据集作为标签数据集，这样对信息重构部进行再训练。
当经由上述训练后，信息重构部的性能达到相应性能要求时，信息重构部的训练结束。网络设备将该训练好的神经网络模型和参数进行存储，并以信息生成部的model ID和版本信息作为该神经网络模型和参数的标志。然后，网络设备通知终端设 备，使能该model ID和版本信息的信息生成部。
当经由上述训练后，信息重构部的性能没有达到要求，网络设备可以配置或调度终端设备上报Y2数据集或Y21数据集。其中Y21数据集是Y2数据集的一部分。当经由上述训练(可继续增加数据集)后，信息重构部的性能达到相应性能要求时，重构部的训练结束。网络设备将该训练好的神经网络模型和参数进行存储，并以信息生成部的model ID和版本信息作为该神经网络模型和参数的标志。网络设备通知终端设备，使能该model ID和版本信息的信息生成部。
如果所有数据集上报并训练后，信息重构部的性能依然达不到要求。网络设备告知终端设备，该model ID和版本不能使用。
在另一个示例中，对于使用双边模型的信号处理模块，网络设备向终端设备进行AI/ML相关能力问询，该问询可包含一个或多个：信号处理模块信息，AI/ML支持信息，AI/ML模型标识信息、版本信息，AI/ML模型数据配置信息，等等。终端设备进行相应的应答汇报。
例如，当网络设备发现终端设备汇报对应某模块具有AI/ML能力，但其具有的信息生成部的model ID或者版本与网络设备的信息重构部的model ID或版本不同时，网络设备利用信息生成部的model ID和版本信息，在其存储器中或者在远端服务器中查找对应该model ID和版本的信息生成部的输入数据集和输出数据集，或者信息重构部的输入数据集和输出数据集，或者共用的两侧数据集。
网络设备基于该输入输出数据集进行训练，当经由训练后，信息重构部的性能达到相应性能要求时，信息重构部的训练结束。网络设备将该训练好的神经网络模型和参数进行存储，并以信息生成部的model ID和版本信息作为该神经网络模型和参数的标志。网络设备可以通知终端设备，使能该model ID和版本信息的信息生成部。
以上各个实施例仅对本申请实施例进行了示例性说明，但本申请不限于此，还可以在以上各个实施例的基础上进行适当的变型。例如，可以单独使用上述各个实施例，也可以将以上各个实施例中的一种或多种结合起来。
由上述实施例可知，第二设备获取被输入信息生成部的第一数据X；以及第二设备获取与所述第一数据X对应并从所述信息生成部输出的第二数据Y。由此，能够以较小的代价获得合适的数据，从而支持双边网络模型架构。
第二方面的实施例
本申请实施例提供一种数据获取方法，从第一设备侧进行说明，与第一方面的实施例相同的内容不再赘述。
图12是本申请实施例的数据获取方法的一示意图，如图12所示，该方法包括：
1201，第一设备获取信息重构部输出的第一数据X；以及
1202，第一设备获取被输入到所述信息重构部的、并且与所述第一数据X对应的第二数据Y。
值得注意的是，以上附图12仅对本申请实施例进行了示意性说明，但本申请不限于此。例如可以适当地调整各个操作之间的执行顺序，此外还可以增加其他的一些操作或者减少其中的某些操作。本领域的技术人员可以根据上述内容进行适当地变型，而不仅限于上述附图12的记载。
在一些实施例中，所述第一设备将所述第一数据X输入到所述信息生成部中，并将所述第二数据Y作为标签数据对所述信息生成部进行训练。
在一些实施例中，所述第二数据Y为输入到所述信息重构部且对应所述第一数据X的数据；所述第一数据X为特定数据集的部分或全部数据，由所述第二数据Y被输入到所述信息重构部后生成。
在一些实施例中，所述第一设备内的所述信息生成部和所述第二设备内的所述信息重构部使用所述第一数据X和所述第二数据Y分别进行训练；所述第一设备内已训练的信息生成部在输入所述第一数据X时，输出为所述第二数据Y或为近似所述第二数据Y的数据；所述第二设备内已训练的信息重构部在输入所述第二数据Y时，输出为所述第一数据X或为近似所述第一数据X的数据。
在一些实施例中，所述第一数据X和所述第二数据Y是成对的数据集；所述成对的数据集具有模型标识信息。
在一些实施例中，一个第一数据X对应多个第二数据Y。
在一些实施例中，所述第一设备根据与所述信息重构部相关的标识信息从所述第一设备内部或从所述第一设备外部获得所述第二数据和/或所述第一数据。
在一些实施例中，所述与信息重构部相关的标识信息包括：所述信息重构部对应的模型标识和/或版本信息，和/或，所述信息重构部对应的模型的数据配置信息。
在一些实施例中，所述第二数据Y通过控制信道或数据信道承载并经由空口传 输，或者，所述第二数据Y根据预定义规则通过数据索引生成，或者第二数据Y的索引由所述第二设备通过空口发送给所述第一设备。
在一些实施例中，所述第一数据X被预先存储在所述第一设备内，或者，所述第一数据X根据预定义规则通过数据索引生成，或者，所述第一数据X由所述第二设备通过空口发送给所述第一设备，或者第一数据X的索引由所述第二设备通过空口发送给所述第一设备。
在一些实施例中，所述AI/ML模型具有模型标识和版本标识；同一AI/ML模型的信息生成部和信息重构部使用相同的模型标识和版本标识，所述信息生成部和所述信息重构部具有不同的子标识。
以下通过示例对如何获取数据进行示意性说明。
例如，发送设备(如网络设备或终端设备)将信息重构部针对输出为{X(数据配置，数据索引|数据)}的输入{Y(数据配置，数据索引|数据)}通过空口发送到接收设备(如终端设备或网络设备)。该数据集可以通过控制信道或数据信道承载。
接收设备的信息生成部利用{X(数据配置，数据索引|数据)}作为其输入数据，对其神经网络模型进行训练，并采用{Y(数据配置，数据索引|数据)}作为label数据，训练的目的是使得在生成部输入x(数据配置，数据索引|数据)时，其中x(数据配置，数据索引|数据)为{X(数据配置，数据索引|数据)}的一个样本，信息生成部的输出等于或近似等于y(数据配置，数据索引|数据)，其中y(数据配置，数据索引|数据)为{Y(数据配置，数据索引|数据)}的一个样本。
对于{X(数据配置，数据索引|数据)}，接收设备获取的方式可以是已经存储在接收设备并通过model ID获取，或者通过model ID从远端的服务器获取，或者，根据预定义的规则由数据索引生成对应的数据。
通过上述方式，可以实现信息重构部不变，信息生成部的模型再训练和更新，如果模型再训练的性能满足要求，在经过或无需经过空口的进一步测试后，信息生成部将对应该再训练和更新的模型标记为该model ID(含版本标识)。
以上仅示意性说明了如何获取数据，更多的示例还可以参考第一方面的实施例。
在一些实施例中，所述第一设备向所述第二设备发送AI/ML相关能力查询；以及所述第一设备接收所述第二设备反馈的AI/ML相关能力响应。其中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识 信息、版本信息、数据配置信息、AI/ML支持训练能力信息、AI/ML升级能力信息。
在一些实施例中，在所述第一设备确定所述第二设备具有AI/ML能力但所述信息生成部的模型标识和/或版本信息与所述信息重构部的模型标识和/或版本信息不同时，所述第一设备获取与所述第二设备的信息重构部的模型标识和/或版本信息对应的所述第一数据和/或所述第二数据。
在一些实施例中，所述信息生成部在利用所述第一数据和所述第二数据完成训练后，所述第一设备将所述信息生成部的模型标识和/或版本信息设置为与所述信息重构部的模型标识和/或版本信息相同。
在一些实施例中，所述第一设备向所述第二设备发送确认其信息重构部可以使用的确认信息,和/或，使能所述信息生成部和/或所述信息重构部的指示信息。
以上仅示意性说明了相关信令过程，更多的示例还可以参考第一方面的实施例。
以上各个实施例仅对本申请实施例进行了示例性说明，但本申请不限于此，还可以在以上各个实施例的基础上进行适当的变型。例如，可以单独使用上述各个实施例，也可以将以上各个实施例中的一种或多种结合起来。
由上述实施例可知，第一设备获取信息重构部输出的第一数据X；以及第一设备获取被输入到所述信息重构部的、并且与所述第一数据X对应的第二数据Y。由此，能够以较小的代价获得合适的数据，从而支持双边网络模型架构。
第三方面的实施例
本申请实施例以CSI生成和上报为例，从网络设备侧进行说明。AI/ML模型包括位于终端设备内的CSI生成部和位于网络设备内的CSI重构部，与第一、二方面的实施例相同的内容不再赘述。
在一些实施例中，共用的成对数据集利用了标准预定义的码书、或者与码书结构类似的DFT矩阵来生成逼近信道矩阵或信道矩阵特征向量，这样很大程度的减少了数据集的开销，特别是如果需要经过空口直接或间接的传递数据集。
图13是本申请实施例的构建数据集的一示例图，示例性示出了利用码书来构建信道特征向量对应数据集。例如，对于Type2和其后续增强版本的码书，码书通过DFT向量的线性组合可以高精度的逼近一个信道矩阵特征向量，利用码书或者码书向量，码书矩阵可以逼近需要反馈的特征向量，或者特征向量矩阵。当共用成对数据 集基于码书构建时，由于基站和终端对码书、或者类似码书的DFT变换矩阵有共识，这样可以通过简单的索引方式来达到码书向量部分在各自侧自己生成的目的，那样，只有被CSI生成部压缩后的比特向量需要在双侧模型部共享，大大减少了开销。
表2
Table 5.2.2.2.1-2:Supported configurations of(N 1,N 2)and(O 1,O 2)
Figure PCTCN2022112299-appb-000002
该表2示意性示出了CSI-RS不同port数的配置信息，对应码书的配置信息。
表3
Table 5.2.2.2.3-5:Codebook for 1-layer and 2-layer CSI reporting using antenna ports 3000to2999+PCSI-RS
Figure PCTCN2022112299-appb-000003
Figure PCTCN2022112299-appb-000004
该表3示意性说明了codebook type2的PMI，通过PMI，即i 1,,q 1,q 2,n 1,n 2,…这样的索引，可以构建码书向量或码书矩阵，其他类型码书的原理类似，可以参考相关技术在此不做进一步说明。
以上对于码书或码书向量等进行了示意性说明，本申请不限于此。
图14是本申请实施例的数据获取方法的一示意图，如图14所示，该方法包括：
1401，网络设备获取被输入所述CSI生成部的特定的码书向量；以及
1402，网络设备获取与所述码书向量对应的从所述CSI生成部输出的CSI比特信息。
值得注意的是，以上附图14仅对本申请实施例进行了示意性说明，但本申请不限于此。例如可以适当地调整各个操作之间的执行顺序，此外还可以增加其他的一些操作或者减少其中的某些操作。本领域的技术人员可以根据上述内容进行适当地变型，而不仅限于上述附图14的记载。
在一些实施例中，所述AI/ML模型为双边模型，具有模型标识和版本标识；所述一个双边模型的CSI生成部和所述CSI重构部具有同样的模型标识和版本标识，所述CSI生成部和所述CSI重构部具有不同的子标识。
例如，包含有CSI生成部AE-x的终端设备和包含有CSI重构部DE-y的网络设备没有经过联合的训练。为了实现基于AI/ML的CSI反馈和恢复，需要在AE-x和DE-y联合工作前，对至少一侧的网络模型进行进一步的训练和调参，以使得不是成对的AE-x和DE-y可以实现满足通信性能要求的联合工作。
在一些实施例中，所述网络设备将所述CSI比特信息输入到所述CSI重构部中，并将所述码书向量作为标签数据对所述CSI重构部进行训练。
在一些实施例中，所述码书向量(X)为输入到所述CSI生成部的数据；所述CSI 比特信息(Y)为将所述码书向量(X)输入到所述CSI生成部后生成的比特序列。
例如，码书向量可以如表4所示：
表4
Table 5.2.2.2.5-5:Codebook for 1-layer.2-layer,3-layer and 4-layer CSI reporting using antenna ports3000to 2999+P CSI-RS
Figure PCTCN2022112299-appb-000005
该表4为eType的码书生成表，同样可以基于该表通过W的各角标索引，生成码书向量或码书矩阵，关于NR各种码书的具体内容可以参照3GPP 38.214。
在一些实施例中，所述CSI比特信息由控制信道或数据信道承载并经由空口传输，和/或，所述码书向量根据预定义规则通过数据索引生成。
例如，基于预定义的预编码码书(precoding codebook)，如Type 1,Type2,etype 2 codebook等NR定义的codebook，可以采用码书所生成的预编码(precoding)矢量集合{V AE-x(i1,i2,…,in)}，作为CSI生成部AE-x(CSI generation part)的AI/ML模型 的输入，{C AE-x(i1,i2,..,in)}为对应的输出比特序列集合。(i1,i2,…,in)为对应码书的索引(index)，或称为PMI，或称为指示序列，具体依据不同类型码书的定义，这里只是示意性说明。
图15是本申请实施例的CSI重构部获取数据的一示例图。
包含有CSI生成部AE-x的终端设备，将生成部输入序列集{V AE-x(x1,x2,…,xn)}对应的序列集合或PMI集合{x1,x2,…,xn}和对应输出序列{C AE-x(x1,x2,..,xn)}发送给含有CSI重构部DE-y的网络设备。DE-y将{C AE-x(x1,x2,..,xn)}作为其输入序列集，因为已收到码书序列集合或PMI集合{x1,x2,..,xn}，DE-y可以根据NR相应码书的生成规则，生成对应的码书向量集合{V AE-x(x1,x2,…,xn)}。
图16是本申请实施例的CSI重构部进行训练的一示例图。
DE-y利用{C AE-x(x1,x2,..,xn)}和{V AE-x(i1,i2,…,in)}作为数据集对其AI/ML model进行训练，训练的目的是使得当输入为C AE-x(x1,x2,..,xn)时，输出矢量逼近V AE-x(i1,i2,…,in)。当训练完成后，就完成了AE-x和DE-y的配对过程。
在一些实施例中，所述网络设备根据与所述CSI生成部相关的标识信息从所述网络设备内部或从所述网络设备外部获得所述CSI比特信息和/或所述码书向量。
在一些实施例中，所述与CSI生成部相关的标识信息包括：所述CSI生成部对应的模型标识和/或版本信息，和/或，所述CSI生成部对应的模型的数据配置信息，和/或，所述CSI生成部的数据索引，和/或，所述CSI生成部的码书向量或DFT向量的指示序列。
例如，当网络设备发现其CSI重构部的model ID或版本与终端设备的CSI生成部的model ID或版本不匹配时，依据CSI生成部的model ID或版本信息去获取相应的输入、输出数据集。可以通过内部存储或通过远端服务器等方式获取对应该model ID及版本的数据集。其中，数据集例如是码书序列集合{x1,x2,…,xn}以及CSI生成部的输出集合{C AE-x(x1,x2,..,xn)}。因为二者都是比特序列，所以很大程度减少了存储量。
基于{x1,x2,…,xn}，对应生成规则可以生成码书向量{V AE-x(x1,x2,…,xn)}。进而可以进行CSI重构部的训练。之后使用CSI生成部的model ID及版本信息标注CSI重构部训练完成时对应的模型参数。
图17是本申请实施例的CSI重构部获取数据的另一示例图。
基于标准定义的码书，如Type 1,Type2,etype 2codebook等NR定义的码书，可以采用码书所生成的precoding矢量集合{V AE-x(i1,i2,…,in)}，作为CSI生成部AE-x(CSI generation part)的AI/ML模型的输入，{C AE-x(i1,i2,..,in)}为对应的输出比特序列集合。(i1,i2,…,in)为对应码书的index序列，或指示序列，或PMI，其排列顺序是预定义或预配置，或配置的。包含有CSI生成部AE-x的终端设备，将输出序列{C AE-x(i1,i2,..,in)}发送给包含有CSI重构部DE-y的网络设备。
图18是本申请实施例的CSI重构部进行训练的另一示例图。图19是本申请实施例的CSI重构部进行训练的另一示例图。
DE-y将{C AE-x(i1,i2,..,in)}作为其输入序列集，因为已知码书索引或标志的顺序(i1,i2,..,in)，DE-y可以据此根据NR相应码书的生成规则，生成对应的码书向量集合{V AE-x(i1,i2,…,in)}。DE-y利用{C AE-x(i1,i2,..,in)}和{V AE-x(i1,i2,…,in)}作为数据集对其AI/ML model进行训练，训练的目的是使得当输入为C AE-x(i1,i2,..,in)时，输出矢量逼近V AE-x(i1,i2,…,in)。当训练完成后，就完成了AE-x和DE-y的配对过程。
在一些实施例中，所述网络设备向所述终端设备发送针对模型标识的模型训练指示和配置信息；所述配置信息包括训练用码书向量或DFT矩阵向量的指示序列；所述网络设备接收所述终端设备发送的用于指示码书向量或DFT向量的指示序列以及所述向量对应的CSI生成部输出比特序列。
在一些实施例中，所述网络设备基于所述指示序列，按照预定义规则生成相应的码书向量或DFT矩阵向量。
在一些实施例中，所述网络设备接收所述终端设备的资源请求；以及所述网络设备为所述终端设备配置上行资源，使得所述终端设备发送所述指示序列和所述CSI生成部输出比特序列。
以上示意性说明了CSI重构部如何获取数据以及如何进行训练，以下再对信令过程进行示意性说明。
在一些实施例中，所述网络设备向所述终端设备发送AI/ML相关能力查询；以及所述网络设备接收所述终端设备反馈的AI/ML相关能力响应。其中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML训练能力信息、AI/ML升级能力信息。
在一些实施例中，在所述网络设备确定所述终端设备具有AI/ML能力但所述CSI生成部的模型标识和/或版本信息与所述CSI重构部的模型标识和/或版本信息不同时，指示或配置所述终端设备发送与所述CSI生成部的模型标识和/或版本信息对应的训练用CSI比特序列和/或对应的码书向量。
在一些实施例中，在所述CSI重构部利用所述CSI比特信息和所述码书向量完成训练后，所述网络设备将所述CSI重构部的模型标识和/或版本信息设置为与所述CSI生成部的模型标识和/或版本信息相同。
在一些实施例中，所述网络设备向所述终端设备发送使能所述CSI生成部和/或所述CSI重构部的指示信息，和/或，发送确认所述CSI生成部可以使用的确认信息。
图20是本申请实施例的CSI重构部训练的一信令示意图。
如图20所示，网络侧让终端侧汇报CSI反馈的AI/ML能力，其中可以包含AI/ML模型ID，版本等model信息的问询。终端侧汇报相关能力信息，其中包含其CSI生成部所用AI/ML模型ID，版本等model信息的指示。网络侧如果具有相应ID和版本的CSI重构部模型，则进入联合工作阶段。
否则，网络侧如果知道终端侧CSI生成部所用AI/ML模型ID和版本与网络侧所具有的一个或多个CSI重构部都不匹配，则说明通信两侧的AI/ML model不能联合工作。网络侧可以配置终端侧上报相应数据，并对CSI重构部进行训练。
网络侧可以问询终端侧CSI生成部输出的比特序列相关的格式信息，终端侧依据预定义的格式汇报其CSI生成部的输出比特信息，例如对应每个码书向量或矩阵的输出信息比特数。
如图20所示，网络侧可以指示终端侧发送对应码书向量集对应的CSI输出集的数据配置信息(如，每个码书向量对应的比特长度)，并发送CSI输出集上报配置。终端侧可以根据配置/指示来上报所对应的CSI输出集。
如图20所示，网络侧使用码书序号对应的码书向量作为label数据，使用码书序号对应的CSI生成器输出作为输入数据，对CSI重构部的model进行训练，使得model输出逼近label数据。当model训练完成后，更新model并更新model标志信息。
如图20所示，网络侧可以配置并指示终端侧采用CSI生成部进行CSI测量及反馈；终端侧根据配置的CSI测量资源，对测量结果使用CSI生成部产生CSI，并根据配置资源进行CSI上报。网络侧可以使用更新的CSI重构部对收到的CSI进行重构。
值得注意的是，以上附图20仅对本申请实施例进行了示意性说明，但本申请不限于此。例如可以适当地调整各个操作之间的执行顺序，此外还可以增加其他的一些操作或者减少其中的某些操作。本领域的技术人员可以根据上述内容进行适当地变型，而不仅限于上述附图20的记载。
对于CSI双边模型，CSI生成部可以基于信道估计后的信道矩阵作为输入，或者以该信道估计后的信道矩阵的特征向量作为输入。可以根据秩(rank)选择输入特征向量的个数，或者基于单layer的特征向量。
同理，对于基于码书向量的数据集，CSI生成部或CSI重构部在获取数据集并进行再训练和进行fine tuning等过程中，需要沟通CSI生成部或CSI重构部的输入输出可能的配置信息，例如输入数据长度、输入数据维度，等等。
在一些实施例中，所述CSI的格式和/或大小被预定义，或者，所述CSI的格式和/或大小由所述网络设备配置。
例如，网络侧需通过问询得知CSI生成部可生成多种CSI比特长度，网络侧可能配置终端侧反馈某一比特长度的CSI生成部的输出，或者多个比特长度，或者所有比特长度的CSI生成部的输出。
例如，表5是终端侧需要反馈的CSI生成部相关信息的例子。
表5
码书指示比特(PMI) 码书向量 CSI生成部比特1 CSI生成部比特2 CSI生成部比特3
A1 A2 A3 A4 A5
B1 B2 B3 B4 B5
在一些实施例中，所述网络设备对所述码书向量进行扩充，并将所述扩充后的码书向量作为所述标签数据。
例如，基于NR已经定义的码书，如Rel-15,16,17,18定义的码书，其所生成的码书向量可能不够完成对AI/ML网络的训练，需要更多的数据集来训练。或者，需要选择更适合CSI生成部和CSI重构部的码书集。
在一些实施例中，所述扩充后的码书向量中，W1的可选波束数目为最多6、8、10、12、16；W2的映射角度为最多5比特，幅度为最多5比特。
例如，可以将现有码书集进行扩展形成扩充码书集或CSI模型训练用码书集。如 目前码书W1最多选L＝4个波束，扩充码书集可以选到L＝6,8,10,12,16。现有的W2映射角度2或3比特，幅度为3比特，可扩展为角度4比特，幅度4比特等。M也可采用比现在标准支持的数量更大的数。
上述CSI模型训练用码书集可以通过配置，得到和目前标准定义的码书(Rel-15、16、17、18等)其中之一的能生成的码书向量的全集，或者根据某一规则的子集。或者也可以通过配置得到扩充的码书集，并依据规则选择子集。或者使用其他的码书索引序列调用上述可能的码书集，调用码书序列的方式是CSI生成部和CSI重构部都知道的。
在一些实施例中，当网络设备发现终端设备具有的模型不一致时，或者CSI生成部进行在线训练和升级，来与CSI重构部匹配，或者CSI重构部进行在线训练和升级，来和CSI生成部匹配。
网络侧可以和终端侧沟通CSI重构部的配置信息。网络侧可以调度终端侧发送CSI生成部确定要使用的码书向量的索引及对应的CSI生成部的输出。该调用可以分批进行；例如先调用一部分数据，训练效果如果不达标就再调度另一部分数据集。通过多次调用来完成训练。
以下再通过示例对本申请进行说明。
例如，基于预定义的码书或DFT变换矩阵(例如，角度时延域二维变换矩阵)可以生成相应的所有可能的码书向量，或DFT变换矩阵对应的所有可能向量。
Figure PCTCN2022112299-appb-000006
例如基于上述Rel-16或Rel-17码书矩阵所生成的向量，其所有可能的码字向量依据预定义顺序排列。DFT变换矩阵对应的向量也可按预定义生成所对应的所有向量。比如，
W＝W 1W 2
Figure PCTCN2022112299-appb-000007
具体采用的DFT基于预定义。可以使用更多的W级联方式，对形成信道的空域，频域，时域特征映射的基函数。
这样对应某一model ID和版本信息的生成部，以及对应某一码书或DFT变换矩阵，并依据所述排列对应的向量作为输入，其输出按照不同输出比特数，存储为数据集。
表6
Figure PCTCN2022112299-appb-000008
上述为某一model ID和版本，CSI生成部向量全集和可能输出的全集。通过索引可以得到子集，子集用于训练。可以多次调用该全集的多个子集进行训练。
子集的调用由索引或索引集的位图(bitmap)指示。位图长度对应表中所有序号长度，从左到右一一对应。位图中的0表示使用该向量，1表示不使用该向量。这里索引集可以对应不同子带，表示不同子带对应的索引。
表7
  Sub band1 Sub band2 Sub band3 Sub band4
PMI-1 0 0 1 1
PMI-2 1 1 0 0
PMI-3 1 1 0 0
PMI-4 1 1 1 0
PMI-5 0 1 1 0
如上表7所示，终端侧或者网络侧针对当前的信道或者历史上经历的信道，可以经过数据筛选和收集，抽取相应的典型信道向量对应的PMI形成数据集，通过bitmap的指示，索引出该信道空域和频域的码书向量对应。
或者，也可以通过DFT变换矩阵对应的DFT基映射的情况来表示。
表8
  时延1 时延2 时延3 时延4
角度-1 1 1 0 0
角度-2 1 1 1 0
角度-3 1 0 0 0
角度-4 1 0 0 0
这里的角度和时延分别由不同的DFT基向量来对应。值得注意的是，以上只是示意性进行了说明，类似的也可以根据配置形成宽带CSI对应的数据集，或者其他方式对应信道特征的数据集。这些数据集可以绑定模型ID，或者根据网络的配置，进一步依据小区模型ID的方式进行绑定，或依据UE模型ID的方式进行绑定，或者依据其他规则来标识。
另一方面，由于该共用成对数据集可以反应信道空域，频域，时域的特征，其标识方式可以相应的体现这些特征，便于在终端或基站侧在能够获取信道特征信息时，应用更适合当前信道条件的数据集来做训练。对应上述方式表示的信道向量数据集，还需要有对应的生成部压缩信息数据集。即所有bitmap为1的信道向量数据集。这样就形成了成对数据集。
其中一个示例是，在得到网络侧的配置后，终端侧将其生成部选择的上述bitmap集和对应的生成部输出集发送给网络侧。当数据集样本较大时，空口开销太大。如终端侧和网络侧可以获取共用成对数据集，则可以通过空口传递该数据集ID的方式来实现空口交互，这样空口开销小。
在一个示例中，基站设备通过监视CSI反馈性能，判断出针对某一model ID(含版本信息)的模型，或者正在使用的CSI模型的性能不好，或者基站基于其他判断决定开始进行CSI重构部的模型训练。
基站配置终端设备针对目前使用的CSI模型ID或某model ID,发送CSI生成部与信道状态信息(CSI)相关的索引标志信息(如bitmap信息)，以及相关训练用码书向量或DFT矩阵向量的指示信息和配置信息。
终端设备基于目前使用的CSI生成部的模型ID或者基站配置的模型ID，以及上述向量指示配置信息，向基站发送该模型对应的CSI生成部与信道状态信息(CSI)相关的索引标志信息(如bitmap信息)，或索引标志信息集。
基站接收该索引信息或索引信息集，根据正在使用的模型ID或其配置的模型ID，获取该模型ID的CSI重构部所需要的输入信息集。基站针对该模型ID(含版本信息)并依据索引或索引集，在其内部或者远端服务器中获取相应的CSI生成部输出向量集，并依据索引或索引集生成相应的CSI生成部输入向量集。将CSI生成部输出向量集作为CSI重构部的输入数据，将CSI生成部输入向量集作为CSI重构部的标签数据，对CSI重构部进行训练。
类似的，基站可针对某一模型ID及版本，发送码书矩阵或DFT矩阵指示及配置信息，以及相应的索引信息或索引信息集给终端设备，终端设备据此从本地或从远端服务器下载相应的CSI生成部(或CSI重构部)输入向量集(输出向量集，码书或DFT变换矩阵对应向量)和输出向量集(输入向量集)。
在另一个示例中，针对model ID和版本，CSI生成部和CSI重构部具有针对不同信道环境的多个数据集供训练使用。数据集以码书向量集，或者DFT变换矩阵(如，角度时延域二维变换矩阵)的生成向量给出，如
Figure PCTCN2022112299-appb-000009
其中各子矩阵的向量基于码书定义或其他预定义给出。这样根据model ID，版本信息，码书或DFT矩阵配置信息，数据集序号，基站和终端设备可以获取CSI重构部或CSI生成部需要的训练数据集。
基站可以配置终端设备发送CSI生成部的上述数据集序号，或者针对model ID版本信息的CSI生成部数据集序号，并基于这些信息获取CSI重构部所需训练数据集。
类似的，基站可以向终端设备发送训练指示，码书或DFT矩阵指示及配置信息，以及CSI重构部的数据集序号，或者针对model ID版本信息的CSI重构部数据集序号。终端设备获取该数据集用于训练，训练完成后发送训练完成信息给基站。另外，其训练完成输出数据集可以用于更新服务器对应数据集。
以上各个实施例仅对本申请实施例进行了示例性说明，但本申请不限于此，还可以在以上各个实施例的基础上进行适当的变型。例如，可以单独使用上述各个实施例，也可以将以上各个实施例中的一种或多种结合起来。
由上述实施例可知，网络设备获取被输入CSI生成部的特定的码书向量；以及网络设备获取与所述码书向量对应的从所述CSI生成部输出的CSI比特信息。由此，能够以较小的代价获得合适的数据，从而支持双边网络模型架构。
第四方面的实施例
本申请实施例以CSI生成和上报为例，从终端设备侧进行说明。AI/ML模型包括位于终端设备内的CSI生成部和位于网络设备内的CSI重构部，与第一至三方面的实施例相同的内容不再赘述。
图21是本申请实施例的数据获取方法的一示意图，如图21所示，该方法包括：
2101，终端设备获取所述CSI重构部输出的码书向量；以及
2102，终端设备获取被输入所述CSI重构部的、并且与所述码书向量对应的CSI比特信息。
值得注意的是，以上附图21仅对本申请实施例进行了示意性说明，但本申请不限于此。例如可以适当地调整各个操作之间的执行顺序，此外还可以增加其他的一些操作或者减少其中的某些操作。本领域的技术人员可以根据上述内容进行适当地变型，而不仅限于上述附图21的记载。
在一些实施例中，所述AI/ML模型为双边模型，具有模型标识和版本标识；同一AI/ML模型的CSI生成部和CSI重构部具有同样的模型标识和版本标识，所述CSI生成部和所述CSI重构部具有不同的子标识。
在一些实施例中，所述CSI比特信息(Y)为输入到所述CSI重构部且对应所述码书向量的信息；所述码书向量(X)为特定数据集的部分或全部数据，由所述CSI比特信息(Y)被输入到所述CSI重构部后生成。
在一些实施例中，所述CSI比特信息由控制信道或数据信道承载并经由空口传输，和/或，所述码书向量根据预定义规则通过数据索引生成。
在一些实施例中，所述终端设备根据与所述CSI重构部相关的标识信息从内存或从所述终端设备内部或从所述终端设备外部获得所述CSI比特信息和/或所述码书向量。
在一些实施例中，所述与CSI重构部相关的标识信息包括：所述CSI重构部对应的模型标识和/或版本信息，和/或，所述重构生成部对应的模型的数据配置信息，和 /或，所述CSI重构部的数据索引，和/或，所述CSI重构部的码书向量或DFT向量的指示序列。
在一些实施例中，所述终端设备接收所述网络设备发送的针对模型标识的模型训练指示和配置信息；所述配置信息包括训练用码书向量或DFT矩阵向量的指示序列以及所述向量对应的CSI重构部输入比特序列。
在一些实施例中，所述终端设备基于所述指示序列，按照预定义规则生成相应的码书向量或DFT矩阵向量。
在一些实施例中，所述终端设备接收所述网络设备的资源配置；以及所述终端设备根据所述资源配置接收所述指示序列和所述CSI重构部输入比特序列。
图22是本申请实施例的CSI生成部获取数据的一示例图。如图22所示，网络侧可以将码书向量对应的CSI数据集发送给终端设备。关于数据或者其他获取方式，可以参考之前的实施例。
在一些实施例中，所述终端设备将所述码书向量输入到所述CSI生成部中，并将所述CSI比特信息作为标签数据对所述CSI生成部进行训练。
图23是本申请实施例的CSI生成部进行训练的一示例图。如图23所示，可以将码书向量序号对应的码书向量集作为输入，将码书向量集对应的CSI数据集作为标签数据，对CSI生成部进行训练。
在一些实施例中，所述终端设备接收所述网络设备发送的AI/ML相关能力查询；以及所述终端设备向所述网络设备反馈AI/ML相关能力响应。其中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML训练能力信息、AI/ML升级能力信息。
在一些实施例中，在所述网络设备确定所述终端设备具有AI/ML能力但所述CSI生成部的模型标识和/或版本信息与所述CSI重构部的模型标识和/或版本信息不同时，指示或配置所述终端设备接收与所述CSI重构部的模型标识和/或版本信息对应的CSI比特信息和/或码书向量。
在一些实施例中，在所述CSI生成部利用所述码书向量和所述CSI比特信息完成训练后，所述终端设备将所述CSI生成部的模型标识和/或版本信息设置为与所述CSI重构部的模型标识和/或版本信息相同。
在一些实施例中，所述终端设备向所述网络设备发送确认其CSI生成部可以使 用的确认信息,和/或，指示所述CSI生成部更新完成的指示信息。
图24是本申请实施例的CSI生成部训练的一信令示意图。
如图24所示，网络侧让终端侧汇报CSI反馈的AI/ML能力，其中可以包含AI/ML模型ID，版本等model信息的问询。终端侧汇报相关能力信息，其中包含其CSI生成部所用AI/ML模型ID，版本等model信息的指示。网络侧如果具有相应ID和版本的CSI重构部模型，则进入联合工作阶段。
否则，网络侧如果知道终端侧CSI生成部所用AI/ML模型ID和版本与网络侧所具有的一个或多个CSI重构部都不匹配，则说明通信两侧的AI/ML model不能联合工作。
网络侧可以指示终端侧进行训练升级，并对CSI生成部进行训练。终端侧可以发送训练升级响应。网络侧可以将CSI重构部对应输出码书向量的输入信息形成数据集。该数据集也可能预存在网络侧。
如图24所示，网络侧可以向终端侧发送码书相关指示信息，码书序列对应的CSI数据集，model标识信息等。
如图24所示，终端侧使用码书序号对应的码书向量作为输入，使用收到的来自网络侧的码书序号对应的CSI数据作为label数据，对CSI生成部进行训练，使得model输出逼近label数据。当model训练完成后，更新model以及model标识信息。然后，可以指示模型更新完成，完成模型配对。
如图24所示，网络侧可以配置并指示终端侧采用CSI生成部进行CSI测量及反馈；终端侧根据配置的CSI测量资源，对测量结果使用更新后的CSI生成部产生CSI，并根据配置资源进行CSI上报。网络侧可以使用CSI重构部对收到的CSI进行重构。
值得注意的是，以上附图24仅对本申请实施例进行了示意性说明，但本申请不限于此。例如可以适当地调整各个操作之间的执行顺序，此外还可以增加其他的一些操作或者减少其中的某些操作。本领域的技术人员可以根据上述内容进行适当地变型，而不仅限于上述附图24的记载。
在一些实施例中，所述CSI的格式和/或大小被预定义，或者，所述CSI的格式和/或大小由所述网络设备配置。
在一些实施例中，所述终端设备对所述码书向量进行扩充，并将所述扩充后的码书向量作为所述标签数据。
在一些实施例中，所述扩充后的码书向量中，W1的可选波束数目为最多6、8、10、12、16；W2的映射角度为最多5比特，幅度为最多5比特。
以下再通过示例对本申请进行说明。
在一个示例中，终端设备接收网络设备发送的模型训练指示信息，和/或，CSI重构部模型数据集指示信息。终端设备获取相应数据集(从本地或远端服务器)作为CSI生成部训练用输出数据集，并依据数据集序号获取或生成CSI生成部输入数据集。
上述训练指示信息可包含模型ID信息、版本信息、码书矩阵或DFT变换矩阵的指示及配置信息、数据集指示信息，指示该model ID和版本相应的数据子集。终端设备可以根据这些数据对CSI生成部进行训练。
如果训练中发现性能不好，终端设备发送训练数据集继续请求信息给网络设备。网络设备继续发送如上的CSI重构部模型数据集指示信息，指示其他的数据集。终端设备获取相应数据集(从本地或远端服务器)，作为CSI生成部训练用输出数据集，并依据数据集序号获取或生成CSI生成部输入数据集。上述过程可继续，直到model训练测性能满足要求，终端设备向网络设备发送训练完成指示信息。
在另一个示例中，网络设备可针对某一模型ID及版本，发送码书矩阵或DFT矩阵指示及配置信息，和相应的索引信息或索引信息集给终端设备，终端设备据此从本地或从远端服务器下载相应的CSI生成部(CSI重构部)输入向量集(输出向量集，码书或DFT变换矩阵对应向量)和输出向量集(输入向量集)。终端设备可以根据这些数据对CSI生成部进行训练。
如果训练性能不能满足要求，终端设备发送训练数据集继续请求信息给网络设备，网络设备发送其他的码书矩阵或DFT矩阵指示及配置信息，和相应的索引信息或索引信息集。终端设备获取相应数据集后继续训练，直到model训练测性能满足要求，终端设备向网络设备发送训练完成指示信息。
以上各个实施例仅对本申请实施例进行了示例性说明，但本申请不限于此，还可以在以上各个实施例的基础上进行适当的变型。例如，可以单独使用上述各个实施例，也可以将以上各个实施例中的一种或多种结合起来。
由上述实施例可知，终端设备获取CSI重构部输出的码书向量；以及终端设备获取被输入所述CSI重构部的、并且与所述码书向量对应的CSI比特信息。由此，能够以较小的代价获得合适的数据，从而支持双边网络模型架构。
第五方面的实施例
本申请实施例提供一种数据获取装置。该装置例如可以是前述的第二设备(终端设备或者网络设备)，也可以是配置于第二设备的某个或某些部件或者组件，与第一至四方面的实施例相同的内容不再赘述。
图25是本申请实施例的数据获取装置的一示意图。AI/ML模型包括位于第一设备内的信息生成部和位于所述第二设备内的信息重构部。如图25所示，数据获取装置2500包括：
第一获取单元2501，其获取被输入所述信息生成部的第一数据X；以及
第二获取单元2502，其获取与所述第一数据X对应并从所述信息生成部输出的第二数据Y。
如图25所示，数据获取装置2500还可以包括：
第一训练单元2503，其将所述第二数据Y输入到所述信息重构部中，并将所述第一数据X作为标签数据对所述信息重构部进行训练。
在一些实施例中，所述第一数据X为输入到所述信息生成部的特定数据集的部分或全部数据；所述第二数据Y为将所述第一数据X输入到所述信息生成部后生成的对应数据。
在一些实施例中，所述第一设备内的所述信息生成部和所述第二设备内的所述信息重构部使用所述第一数据X和所述第二数据Y分别进行训练；
所述第一设备内已训练的信息生成部在输入所述第一数据X时，输出为所述第二数据Y或为近似所述第二数据Y的数据；所述第二设备内已训练的信息重构部在输入所述第二数据Y时，输出为所述第一数据X或为近似所述第一数据X的数据。
在一些实施例中，所述第一数据X和所述第二数据Y是成对的数据集；所述成对的数据集具有模型标识信息。
在一些实施例中，一个第一数据X对应多个第二数据Y。
在一些实施例中，所述第一获取单元2501或所述第二获取单元2502根据与所述信息生成部相关的标识信息从所述第二设备内部或从所述第二设备外部获得所述第一数据X和/或所述第二数据Y。
在一些实施例中，与所述信息生成部相关的标识信息包括：所述信息生成部对应 的模型标识和/或版本信息，和/或，所述信息生成部对应的模型的数据配置信息。
在一些实施例中，所述第二数据Y通过控制信道或数据信道承载并经由空口传输，或者，所述第二数据Y根据预定义规则通过数据索引生成，或者第二数据Y的索引由所述第一设备通过空口发送给所述第二设备。
在一些实施例中，所述第一数据X被预先存储在所述第二设备内，或者，所述第一数据X根据预定义规则通过数据索引生成，或者，所述第一数据X由所述第一设备通过空口发送给所述第二设备，或者第一数据X的索引由所述第一设备通过空口发送给所述第二设备。
在一些实施例中，所述AI/ML模型具有模型标识和版本标识；同一AI/ML模型的信息生成部和信息重构部使用相同的模型标识和版本标识，所述信息生成部和所述信息重构部具有不同的子标识。
在一些实施例中，所述装置还包括：
发送单元，其向所述第一设备发送AI/ML相关能力查询；以及
接收单元，其接收所述第一设备反馈的AI/ML相关能力响应。
在一些实施例中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML支持训练能力信息、AI/ML升级能力信息。
在一些实施例中，在所述第二设备确定所述第一设备具有AI/ML能力但所述信息生成部的模型标识和/或版本信息与所述信息重构部的模型标识和/或版本信息不同时，所述第一获取单元或所述第二获取单元获取与所述第一设备的信息生成部的模型标识和/或版本信息对应的所述第一数据和/或所述第二数据。
在一些实施例中，所述信息重构部在利用所述第一数据和所述第二数据完成训练后，所述信息重构部的模型标识和/或版本信息被设置为与所述信息生成部的模型标识和/或版本信息相同。
在一些实施例中，所述发送单元还向所述第一设备发送确认其信息生成部可以使用的确认信息,和/或，发送使能所述信息生成部和/或所述信息重构部的指示信息。
以下再以CSI为例进行说明。第二设备为网络设备，AI/ML模型包括位于终端设备内的CSI生成部和位于网络设备内的CSI重构部。
在一些实施例中，第一获取单元2501获取被输入所述CSI生成部的特定的码书 向量；以及第二获取单元2502获取与所述码书向量对应的从所述CSI生成部输出的CSI比特信息。
在一些实施例中，第一训练单元2503将所述CSI比特信息输入到所述CSI重构部中，并将所述码书向量作为标签数据对所述CSI重构部进行训练。
在一些实施例中，所述码书向量(X)为输入到所述CSI生成部的数据；所述CSI比特信息(Y)为将所述码书向量(X)输入到所述CSI生成部后生成的比特序列。
在一些实施例中，所述CSI比特信息由控制信道或数据信道承载并经由空口传输，和/或，所述码书向量根据预定义规则通过数据索引生成。
在一些实施例中，所述网络设备根据与所述CSI生成部相关的标识信息从所述网络设备内部或从所述网络设备外部获得所述CSI比特信息和/或所述码书向量。
在一些实施例中，所述与CSI生成部相关的标识信息包括：所述CSI生成部对应的模型标识和/或版本信息，和/或，所述CSI生成部对应的模型的数据配置信息，和/或，所述CSI生成部的数据索引，和/或，所述CSI生成部的码书向量或DFT向量的指示序列。
在一些实施例中，所述网络设备向所述终端设备发送针对模型标识的模型训练指示和配置信息；所述配置信息包括训练用码书向量或DFT矩阵向量的指示序列；
所述网络设备接收所述终端设备发送的用于指示码书向量或DFT向量的指示序列以及所述向量对应的CSI生成部输出比特序列。
在一些实施例中，所述网络设备基于所述指示序列，按照预定义规则生成相应的码书向量或DFT矩阵向量。
在一些实施例中，所述网络设备接收所述终端设备的资源请求；以及所述网络设备为所述终端设备配置上行资源，使得所述终端设备发送所述指示序列和所述CSI生成部输出比特序列。
在一些实施例中，所述AI/ML模型为双边模型，具有模型标识和版本标识；所述一个双边模型的CSI生成部和所述CSI重构部具有同样的模型标识和版本标识，所述CSI生成部和所述CSI重构部具有不同的子标识。
在一些实施例中，所述网络设备向所述终端设备发送AI/ML相关能力查询；以及所述网络设备接收所述终端设备反馈的AI/ML相关能力响应。
在一些实施例中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、 AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML训练能力信息、AI/ML升级能力信息。
在一些实施例中，在所述网络设备确定所述终端设备具有AI/ML能力但所述CSI生成部的模型标识和/或版本信息与所述CSI重构部的模型标识和/或版本信息不同时，指示或配置所述终端设备发送与所述CSI生成部的模型标识和/或版本信息对应的训练用CSI比特序列和/或对应的码书向量。
在一些实施例中，所述CSI的格式和/或大小被预定义，或者，所述CSI的格式和/或大小由所述网络设备配置。
在一些实施例中，在所述CSI重构部利用所述CSI比特信息和所述码书向量完成训练后，所述网络设备将所述CSI重构部的模型标识和/或版本信息设置为与所述CSI生成部的模型标识和/或版本信息相同。
在一些实施例中，所述网络设备向所述终端设备发送使能所述CSI生成部和/或所述CSI重构部的指示信息，和/或，发送确认所述CSI生成部可以使用的确认信息。
在一些实施例中，所述网络设备对所述码书向量进行扩充，并将所述扩充后的码书向量作为所述标签数据。
在一些实施例中，所述扩充后的码书向量中，W1的可选波束数目为最多6、8、10、12、16；W2的映射角度为最多5比特，幅度为最多5比特。
以上各个实施例仅对本申请实施例进行了示例性说明，但本申请不限于此，还可以在以上各个实施例的基础上进行适当的变型。例如，可以单独使用上述各个实施例，也可以将以上各个实施例中的一种或多种结合起来。
值得注意的是，以上仅对与本申请相关的各部件或模块进行了说明，但本申请不限于此。数据获取装置2500还可以包括其他部件或者模块，关于这些部件或者模块的具体内容，可以参考相关技术。
此外，为了简单起见，图25中仅示例性示出了各个部件或模块之间的连接关系或信号走向，但是本领域技术人员应该清楚的是，可以采用总线连接等各种相关技术。上述各个部件或模块可以通过例如处理器、存储器、发射机、接收机等硬件设施来实现；本申请实施并不对此进行限制。
由上述实施例可知，第二设备获取被输入信息生成部的第一数据X；以及第二设备获取与所述第一数据X对应并从所述信息生成部输出的第二数据Y。由此，能够 以较小的代价获得合适的数据，从而支持双边网络模型架构。
第六方面的实施例
本申请实施例提供一种数据获取装置。该装置例如可以是前述的第一设备(终端设备或网络设备)，也可以是配置于第一设备的某个或某些部件或者组件，与第一至四方面的实施例相同的内容不再赘述。
图26是本申请实施例的数据获取装置的一示意图。AI/ML模型包括位于所述第一设备内的信息生成部和位于第二设备内的信息重构部，如图26所示，数据获取装置2600包括：
第三获取单元2601，其获取所述信息重构部输出的第一数据X；以及
第四获取单元2602，其获取被输入到所述信息重构部的、并且与所述第一数据X对应的第二数据Y。
如图26所示，数据获取装置2600还可以包括：
第二训练单元2603，其将所述第一数据X输入到所述信息生成部中，并将所述第二数据Y作为标签数据对所述信息生成部进行训练。
在一些实施例中，所述第二数据Y为输入到所述信息重构部且对应所述第一数据X的数据；所述第一数据X为特定数据集的部分或全部数据，由所述第二数据Y被输入到所述信息重构部后生成。
在一些实施例中，所述第一设备内的所述信息生成部和所述第二设备内的所述信息重构部使用所述第一数据X和所述第二数据Y分别进行训练；
所述第一设备内已训练的信息生成部在输入所述第一数据X时，输出为所述第二数据Y或为近似所述第二数据Y的数据；所述第二设备内已训练的信息重构部在输入所述第二数据Y时，输出为所述第一数据X或为近似所述第一数据X的数据。
在一些实施例中，所述第一数据X和所述第二数据Y是成对的数据集；所述成对的数据集具有模型标识信息。
在一些实施例中，一个所述第一数据X对应多个所述第二数据Y。
在一些实施例中，所述第一设备根据与所述信息重构部相关的标识信息从所述第一设备内部或从所述第一设备外部获得所述第二数据和/或所述第一数据。
在一些实施例中，所述与信息重构部相关的标识信息包括：所述信息重构部对应 的模型标识和/或版本信息，和/或，所述信息重构部对应的模型的数据配置信息。
在一些实施例中，所述第二数据Y通过控制信道或数据信道承载并经由空口传输，或者，所述第二数据Y根据预定义规则通过数据索引生成，或者第二数据Y的索引由所述第二设备通过空口发送给所述第一设备。
在一些实施例中，所述第一数据X被预先存储在所述第一设备内，或者，所述第一数据X根据预定义规则通过数据索引生成，或者，所述第一数据X由所述第二设备通过空口发送给所述第一设备，或者第一数据X的索引由所述第二设备通过空口发送给所述第一设备。
在一些实施例中，所述AI/ML模型具有模型标识和版本标识；同一AI/ML模型的信息生成部和信息重构部使用相同的模型标识和版本标识，所述信息生成部和所述信息重构部具有不同的子标识。
在一些实施例中，所述装置还包括：
发送单元，其向所述第二设备发送AI/ML相关能力查询；以及
接收单元，其接收所述第二设备反馈的AI/ML相关能力响应。
在一些实施例中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML支持训练能力信息、AI/ML升级能力信息。
在一些实施例中，在所述第一设备确定所述第二设备具有AI/ML能力但所述信息生成部的模型标识和/或版本信息与所述信息重构部的模型标识和/或版本信息不同时，所述第一设备获取与所述第二设备的信息重构部的模型标识和/或版本信息对应的所述第一数据和/或所述第二数据。
在一些实施例中，所述信息生成部在利用所述第一数据和所述第二数据完成训练后，所述第一设备将所述信息生成部的模型标识和/或版本信息设置为与所述信息重构部的模型标识和/或版本信息相同。
在一些实施例中，所述发送单元向所述第二设备发送确认其信息重构部可以使用的确认信息,和/或，使能所述信息生成部和/或所述信息重构部的指示信息。
以下再以CSI为例进行说明，第一设备为终端设备。AI/ML模型包括位于终端设备内的CSI生成部和位于网络设备内的CSI重构部。
第三获取单元2601获取所述CSI重构部输出的码书向量；以及第四获取单元 2602获取被输入所述CSI重构部的、并且与所述码书向量对应的CSI比特信息。
在一些实施例中，第二训练单元2603将所述码书向量输入到所述CSI生成部中，并将所述CSI比特信息作为标签数据对所述CSI生成部进行训练。
在一些实施例中，所述CSI比特信息(Y)为输入到所述CSI重构部且对应所述码书向量的信息；所述码书向量(X)为特定数据集的部分或全部数据，由所述CSI比特信息(Y)被输入到所述CSI重构部后生成。
在一些实施例中，所述CSI比特信息由控制信道或数据信道承载并经由空口传输，和/或，所述码书向量根据预定义规则通过数据索引生成。
在一些实施例中，所述终端设备根据与所述CSI重构部相关的标识信息从内存或从所述终端设备内部或从所述终端设备外部获得所述CSI比特信息和/或所述码书向量。
在一些实施例中，所述与CSI重构部相关的标识信息包括：所述CSI重构部对应的模型标识和/或版本信息，和/或，所述重构生成部对应的模型的数据配置信息，和/或，所述CSI重构部的数据索引，和/或，所述CSI重构部的码书向量或DFT向量的指示序列。
在一些实施例中，所述终端设备接收所述网络设备发送的针对模型标识的模型训练指示和配置信息；所述配置信息包括训练用码书向量或DFT矩阵向量的指示序列以及所述向量对应的CSI重构部输入比特序列。
在一些实施例中，所述终端设备基于所述指示序列，按照预定义规则生成相应的码书向量或DFT矩阵向量。
在一些实施例中，所述终端设备接收所述网络设备的资源配置；以及所述终端设备根据所述资源配置接收所述指示序列和所述CSI重构部输入比特序列。
在一些实施例中，所述AI/ML模型为双边模型，具有模型标识和版本标识；同一AI/ML模型的CSI生成部和CSI重构部具有同样的模型标识和版本标识，所述CSI生成部和所述CSI重构部具有不同的子标识。
在一些实施例中，所述终端设备接收所述网络设备发送的AI/ML相关能力查询；以及所述终端设备向所述网络设备反馈AI/ML相关能力响应。
在一些实施例中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML训练能力 信息、AI/ML升级能力信息。
在一些实施例中，在所述网络设备确定所述终端设备具有AI/ML能力但所述CSI生成部的模型标识和/或版本信息与所述CSI重构部的模型标识和/或版本信息不同时，指示或配置所述终端设备接收与所述CSI重构部的模型标识和/或版本信息对应的CSI比特信息和/或码书向量。
在一些实施例中，所述CSI的格式和/或大小被预定义，或者，所述CSI的格式和/或大小由所述网络设备配置。
在一些实施例中，在所述CSI生成部利用所述码书向量和所述CSI比特信息完成训练后，所述终端设备将所述CSI生成部的模型标识和/或版本信息设置为与所述CSI重构部的模型标识和/或版本信息相同。
在一些实施例中，所述终端设备向所述网络设备发送确认其CSI生成部可以使用的确认信息,和/或，指示所述CSI生成部更新完成的指示信息。
在一些实施例中，所述终端设备对所述码书向量进行扩充，并将所述扩充后的码书向量作为所述标签数据。
在一些实施例中，所述扩充后的码书向量中，W1的可选波束数目为最多6、8、10、12、16；W2的映射角度为最多5比特，幅度为最多5比特。
以上各个实施例仅对本申请实施例进行了示例性说明，但本申请不限于此，还可以在以上各个实施例的基础上进行适当的变型。例如，可以单独使用上述各个实施例，也可以将以上各个实施例中的一种或多种结合起来。
值得注意的是，以上仅对与本申请相关的各部件或模块进行了说明，但本申请不限于此。数据获取装置2600还可以包括其他部件或者模块，关于这些部件或者模块的具体内容，可以参考相关技术。
此外，为了简单起见，图26中仅示例性示出了各个部件或模块之间的连接关系或信号走向，但是本领域技术人员应该清楚的是，可以采用总线连接等各种相关技术。上述各个部件或模块可以通过例如处理器、存储器、发射机、接收机等硬件设施来实现；本申请实施并不对此进行限制。
由上述实施例可知，第一设备获取信息重构部输出的第一数据X；以及第一设备获取被输入到所述信息重构部的、并且与所述第一数据X对应的第二数据Y。由此，能够以较小的代价获得合适的数据，从而支持双边网络模型架构。
第七方面的实施例
本申请实施例还提供一种通信系统，可以参考图1，与第一至六方面的实施例相同的内容不再赘述。
在一些实施例中，AI/ML模型包括位于第一设备内的信息生成部和位于第二设备内的信息重构部，通信系统100至少可以包括：
第一设备，其获取所述信息重构部输出的第一数据X；以及获取被输入到所述信息重构部的、并且与所述第一数据X对应的第二数据Y；
第二设备，其获取被输入所述信息生成部的第一数据X；以及获取与所述第一数据X对应并从所述信息生成部输出的第二数据Y。
在一些实施例中，AI/ML模型包括位于终端设备内的CSI生成部和位于网络设备内的CSI重构部；通信系统100至少可以包括：
网络设备，其获取被输入所述CSI生成部的特定的码书向量；以及获取与所述码书向量对应的从所述CSI生成部输出的CSI比特信息；
终端设备，其获取所述CSI重构部输出的码书向量；以及获取被输入所述CSI重构部的、并且与所述码书向量对应的CSI比特信息。
本申请实施例还提供一种网络设备，例如可以是基站，但本申请不限于此，还可以是其他的网络设备。
图27是本申请实施例的网络设备的构成示意图。如图27所示，网络设备2700可以包括：处理器2710(例如中央处理器CPU)和存储器2720；存储器2720耦合到处理器2710。其中该存储器2720可存储各种数据；此外还存储信息处理的程序2730，并且在处理器2710的控制下执行该程序2730。
例如，处理器2710可以被配置为执行程序而实现如第一方面的实施例所述的数据获取方法。例如处理器2710可以被配置为进行如下的控制：获取被输入信息生成部的第一数据X；以及获取与所述第一数据X对应并从信息生成部输出的第二数据Y。
再例如，处理器2710可以被配置为执行程序而实现如第二方面的实施例所述的数据获取方法。例如处理器2710可以被配置为进行如下的控制：获取信息重构部输出的第一数据X；以及获取被输入到所述信息重构部的、并且与所述第一数据X对 应的第二数据Y。
再例如，处理器2710可以被配置为执行程序而实现如第三方面的实施例所述的数据获取方法。例如处理器2710可以被配置为进行如下的控制：获取被输入CSI生成部的特定的码书向量；以及获取与所述码书向量对应的从所述CSI生成部输出的CSI比特信息。
此外，如图27所示，网络设备2700还可以包括：收发机2740和天线2750等；其中，上述部件的功能与现有技术类似，此处不再赘述。值得注意的是，网络设备2700也并不是必须要包括图27中所示的所有部件；此外，网络设备2700还可以包括图27中没有示出的部件，可以参考现有技术。
本申请实施例还提供一种终端设备，但本申请不限于此，还可以是其他的设备。
图28是本申请实施例的终端设备的示意图。如图28所示，该终端设备2800可以包括处理器2810和存储器2820；存储器2820存储有数据和程序，并耦合到处理器2810。值得注意的是，该图是示例性的；还可以使用其他类型的结构，来补充或代替该结构，以实现电信功能或其他功能。
例如，处理器2810可以被配置为执行程序而实现如第一方面的实施例所述的数据获取方法。例如处理器2810可以被配置为进行如下的控制：获取被输入信息生成部的第一数据X；以及获取与所述第一数据X对应并从信息生成部输出的第二数据Y。
再例如，处理器2810可以被配置为执行程序而实现如第二方面的实施例所述的数据获取方法。例如处理器2810可以被配置为进行如下的控制：获取信息重构部输出的第一数据X；以及获取被输入到所述信息重构部的、并且与所述第一数据X对应的第二数据Y。
再例如，处理器2810可以被配置为执行程序而实现如第四方面的实施例所述的数据获取方法。例如处理器2810可以被配置为进行如下的控制：获取CSI重构部输出的码书向量；以及获取被输入所述CSI重构部的、并且与所述码书向量对应的CSI比特信息。
如图28所示，该终端设备2800还可以包括：通信模块2830、输入单元2840、显示器2850、电源2860。其中，上述部件的功能与现有技术类似，此处不再赘述。值得注意的是，终端设备2800也并不是必须要包括图28中所示的所有部件，上述部 件并不是必需的；此外，终端设备2800还可以包括图28中没有示出的部件，可以参考现有技术。
本申请实施例还提供一种计算机程序，其中当在终端设备中执行所述程序时，所述程序使得所述终端设备执行第一、二、四方面的实施例所述的数据获取方法。
本申请实施例还提供一种存储有计算机程序的存储介质，其中所述计算机程序使得终端设备执行第一、二、四方面的实施例所述的数据获取方法。
本申请实施例还提供一种计算机程序，其中当在网络设备中执行所述程序时，所述程序使得所述网络设备执行第一、二、三方面的实施例所述的数据获取方法。
本申请实施例还提供一种存储有计算机程序的存储介质，其中所述计算机程序使得网络设备执行第一、二、三方面的实施例所述的数据获取方法。
本申请以上的装置和方法可以由硬件实现，也可以由硬件结合软件实现。本申请涉及这样的计算机可读程序，当该程序被逻辑部件所执行时，能够使该逻辑部件实现上文所述的装置或构成部件，或使该逻辑部件实现上文所述的各种方法或步骤。本申请还涉及用于存储以上程序的存储介质，如硬盘、磁盘、光盘、DVD、flash存储器等。
结合本申请实施例描述的方法/装置可直接体现为硬件、由处理器执行的软件模块或二者组合。例如，图中所示的功能框图中的一个或多个和/或功能框图的一个或多个组合，既可以对应于计算机程序流程的各个软件模块，亦可以对应于各个硬件模块。这些软件模块，可以分别对应于图中所示的各个步骤。这些硬件模块例如可利用现场可编程门阵列(FPGA)将这些软件模块固化而实现。
软件模块可以位于RAM存储器、闪存、ROM存储器、EPROM存储器、EEPROM存储器、寄存器、硬盘、移动磁盘、CD-ROM或者本领域已知的任何其它形式的存储介质。可以将一种存储介质耦接至处理器，从而使处理器能够从该存储介质读取信息，且可向该存储介质写入信息；或者该存储介质可以是处理器的组成部分。处理器和存储介质可以位于ASIC中。该软件模块可以存储在移动终端的存储器中，也可以存储在可插入移动终端的存储卡中。例如，若设备(如移动终端)采用的是较大容量的MEGA-SIM卡或者大容量的闪存装置，则该软件模块可存储在该MEGA-SIM卡或者大容量的闪存装置中。
针对附图中描述的功能方框中的一个或多个和/或功能方框的一个或多个组合，可以实现为用于执行本申请所描述功能的通用处理器、数字信号处理器(DSP)、专用 集成电路(ASIC)、现场可编程门阵列(FPGA)或者其它可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件或者其任意适当组合。针对附图描述的功能方框中的一个或多个和/或功能方框的一个或多个组合，还可以实现为计算设备的组合，例如，DSP和微处理器的组合、多个微处理器、与DSP通信结合的一个或多个微处理器或者任何其它这种配置。
以上结合具体的实施方式对本申请进行了描述，但本领域技术人员应该清楚，这些描述都是示例性的，并不是对本申请保护范围的限制。本领域技术人员可以根据本申请的精神和原理对本申请做出各种变型和修改，这些变型和修改也在本申请的范围内。
关于包括以上实施例的实施方式，还公开下述的附记：
1.一种数据获取方法，其中，AI/ML模型包括位于第一设备内的信息生成部和位于第二设备内的信息重构部，所述方法包括：
所述第二设备获取被输入所述信息生成部的第一数据X；以及
所述第二设备获取与所述第一数据X对应并从所述信息生成部输出的第二数据Y。
2.根据附记1所述的方法，其中，所述方法还包括：
所述第二设备将所述第二数据Y输入到所述信息重构部中，并将所述第一数据X作为标签数据对所述信息重构部进行训练。
3.根据附记1或2所述的方法，其中，所述第一数据X为输入到所述信息生成部的特定数据集的部分或全部数据；所述第二数据Y为将所述第一数据X输入到所述信息生成部后生成的对应数据。
4.根据附记1至3任一项所述的方法，其中，所述第一设备内的所述信息生成部和所述第二设备内的所述信息重构部使用所述第一数据X和所述第二数据Y分别进行训练；
所述第一设备内已训练的信息生成部在输入所述第一数据X时，输出为所述第二数据Y或为近似所述第二数据Y的数据；
所述第二设备内已训练的信息重构部在输入所述第二数据Y时，输出为所述第一数据X或为近似所述第一数据X的数据。
5.根据附记1至4任一项所述的方法，其中，所述第一数据X和所述第二数据 Y是成对的数据集；所述成对的数据集具有模型标识信息。
6.根据附记5所述的方法，其中，一个第一数据X对应多个第二数据Y。
7.根据附记6所述的方法，其中，所述第二数据的格式被预定义或者被配置。
8.根据附记4至7任一项所述的方法，其中，已训练的所述成对的数据集被注册模型ID和/或版本ID。
9.根据附记8所述的方法，其中，所述模型ID和/或版本ID给上传到核心网、外部服务器、或网络设备中。
10.根据附记1至9任一项所述的方法，其中，
所述第二设备根据与所述信息生成部相关的标识信息从所述第二设备内部或从所述第二设备外部获得所述第一数据X和/或所述第二数据Y。
11.根据附记10所述的方法，其中，所述与信息生成部相关的标识信息包括：所述信息生成部对应的模型标识和/或版本信息，和/或，所述信息生成部对应的模型的数据配置信息。
12.根据附记1至11任一项所述的方法，其中，所述第二数据Y通过控制信道或数据信道承载并经由空口传输，或者，所述第二数据Y根据预定义规则通过数据索引生成，或者第二数据Y的索引由所述第一设备通过空口发送给所述第二设备。
13.根据附记1至12任一项所述的方法，其中，所述第一数据X被预先存储在所述第二设备内，或者，所述第一数据X根据预定义规则通过数据索引生成，或者，所述第一数据X由所述第一设备通过空口发送给所述第二设备，或者第一数据X的索引由所述第一设备通过空口发送给所述第二设备。
14.根据附记1至13任一项所述的方法，其中，所述AI/ML模型具有模型标识和版本标识；同一AI/ML模型的信息生成部和信息重构部使用相同的模型标识和版本标识，所述信息生成部和所述信息重构部具有不同的子标识。
15.根据附记1至14任一项所述的方法，其中，所述方法还包括：
所述第二设备向所述第一设备发送AI/ML相关能力查询；以及
所述第二设备接收所述第一设备反馈的AI/ML相关能力响应。
16.根据附记15所述的方法，其中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML支持训练能力信息、AI/ML升级能力信息。
17.根据附记15所述的方法，其中，所述方法还包括：
在所述第二设备确定所述第一设备具有AI/ML能力但所述信息生成部的模型标识和/或版本信息与所述信息重构部的模型标识和/或版本信息不同时，所述第二设备获取与所述第一设备的信息生成部的模型标识和/或版本信息对应的所述第一数据和/或所述第二数据。
18.根据附记17所述的方法，其中，所述方法还包括：
所述信息重构部在利用所述第一数据和所述第二数据完成训练后，所述第二设备将所述信息重构部的模型标识和/或版本信息设置为与所述信息生成部的模型标识和/或版本信息相同。
19.根据附记18所述的方法，其中，所述方法还包括：
所述第二设备向所述第一设备发送确认其信息生成部可以使用的确认信息,和/或，发送使能所述信息生成部和/或所述信息重构部的指示信息。
20.一种数据获取方法，其中，AI/ML模型包括位于第一设备内的信息生成部和位于第二设备内的信息重构部，所述方法包括：
所述第一设备获取所述信息重构部输出的第一数据X；以及
所述第一设备获取被输入到所述信息重构部的、并且与所述第一数据X对应的第二数据Y。
21.根据附记20所述的方法，其中，所述方法还包括：
所述第一设备将所述第一数据X输入到所述信息生成部中，并将所述第二数据Y作为标签数据对所述信息生成部进行训练。
22.根据附记20或21所述的方法，其中，所述第二数据Y为输入到所述信息重构部且对应所述第一数据X的数据；所述第一数据X为特定数据集的部分或全部数据，由所述第二数据Y被输入到所述信息重构部后生成。
23.根据附记20至22任一项所述的方法，其中，所述第一设备内的所述信息生成部和所述第二设备内的所述信息重构部使用所述第一数据X和所述第二数据Y分别进行训练；
所述第一设备内已训练的信息生成部在输入所述第一数据X时，输出为所述第二数据Y或为近似所述第二数据Y的数据；
所述第二设备内已训练的信息重构部在输入所述第二数据Y时，输出为所述第 一数据X或为近似所述第一数据X的数据。
24.根据附记20至23任一项所述的方法，其中，所述第一数据X和所述第二数据Y是成对的数据集；所述成对的数据集具有模型标识信息。
25.根据附记24所述的方法，其中，一个第一数据X对应多个第二数据Y。
26.根据附记25所述的方法，其中，所述第二数据的格式被预定义或者被配置。
27.根据附记23至25任一项所述的方法，其中，已训练的所述成对的数据集被注册模型ID和/或版本ID。
28.根据附记27所述的方法，其中，所述模型ID和/或版本ID给上传到核心网、外部服务器、或网络设备中。
29.根据附记20至28任一项所述的方法，其中，
所述第一设备根据与所述信息重构部相关的标识信息从所述第一设备内部或从所述第一设备外部获得所述第二数据和/或所述第一数据。
30.根据附记29所述的方法，其中，所述与信息重构部相关的标识信息包括：所述信息重构部对应的模型标识和/或版本信息，和/或，所述信息重构部对应的模型的数据配置信息。
31.根据附记20至30任一项所述的方法，其中，所述第二数据Y通过控制信道或数据信道承载并经由空口传输，或者，所述第二数据Y根据预定义规则通过数据索引生成，或者第二数据Y的索引由所述第二设备通过空口发送给所述第一设备。
32.根据附记20至31任一项所述的方法，其中，所述第一数据X被预先存储在所述第一设备内，或者，所述第一数据X根据预定义规则通过数据索引生成，或者，所述第一数据X由所述第二设备通过空口发送给所述第一设备，或者第一数据X的索引由所述第二设备通过空口发送给所述第一设备。
33.根据附记20至32任一项所述的方法，其中，所述AI/ML模型具有模型标识和版本标识；同一AI/ML模型的信息生成部和信息重构部使用相同的模型标识和版本标识，所述信息生成部和所述信息重构部具有不同的子标识。
34.根据附记20至33任一项所述的方法，其中，所述方法还包括：
所述第一设备向所述第二设备发送AI/ML相关能力查询；以及
所述第一设备接收所述第二设备反馈的AI/ML相关能力响应。
35.根据附记34所述的方法，其中，所述AI/ML相关能力包括如下至少之一： 信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML支持训练能力信息、AI/ML升级能力信息。
36.根据附记34所述的方法，其中，所述方法还包括：
在所述第一设备确定所述第二设备具有AI/ML能力但所述信息生成部的模型标识和/或版本信息与所述信息重构部的模型标识和/或版本信息不同时，所述第一设备获取与所述第二设备的信息重构部的模型标识和/或版本信息对应的所述第一数据和/或所述第二数据。
37.根据附记36所述的方法，其中，所述方法还包括：
所述信息生成部在利用所述第一数据和所述第二数据完成训练后，所述第一设备将所述信息生成部的模型标识和/或版本信息设置为与所述信息重构部的模型标识和/或版本信息相同。
38.根据附记37所述的方法，其中，所述方法还包括：
所述第一设备向所述第二设备发送确认其信息重构部可以使用的确认信息,和/或，使能所述信息生成部和/或所述信息重构部的指示信息。
39.一种数据获取方法，其中，AI/ML模型包括位于终端设备内的CSI生成部和位于网络设备内的CSI重构部，所述方法包括：
所述网络设备获取被输入所述CSI生成部的特定的码书向量；以及
所述网络设备获取与所述码书向量对应的从所述CSI生成部输出的CSI比特信息。
40.根据附记39所述的方法，其中，所述方法还包括：
所述网络设备将所述CSI比特信息输入到所述CSI重构部中，并将所述码书向量作为标签数据对所述CSI重构部进行训练。
41.根据附记39或40所述的方法，其中，所述码书向量(X)为输入到所述CSI生成部的数据；所述CSI比特信息(Y)为将所述码书向量(X)输入到所述CSI生成部后生成的比特序列。
42.根据附记39至41任一项所述的方法，其中，所述CSI比特信息由控制信道或数据信道承载并经由空口传输，和/或，所述码书向量根据预定义规则通过数据索引或指示序列或预编码矩阵指示(PMI)生成。
43.根据附记39至41任一项所述的方法，其中，
所述网络设备根据与所述CSI生成部相关的标识信息从所述网络设备内部或从所述网络设备外部获得所述CSI比特信息和/或所述码书向量。
44.根据附记43所述的方法，其中，所述与CSI生成部相关的标识信息包括：所述CSI生成部对应的模型标识和/或版本信息，和/或，所述CSI生成部对应的模型的数据配置信息，和/或，所述CSI生成部的数据索引，和/或，所述CSI生成部的码书向量或DFT向量的指示序列。
45.根据附记39至44任一项所述的方法，其中，所述方法还包括：
所述网络设备向所述终端设备发送针对模型标识的模型训练指示和配置信息；所述配置信息包括训练用码书向量或DFT矩阵向量的指示序列；
所述网络设备接收所述终端设备发送的用于指示码书向量或DFT向量的指示序列以及所述向量对应的CSI生成部输出比特序列。
46.根据附记45所述的方法，其中，所述方法还包括：
所述网络设备基于所述指示序列，按照预定义规则生成相应的码书向量或DFT矩阵向量。
47.根据附记45或46所述的方法，其中，所述方法还包括：
所述网络设备接收所述终端设备的资源请求；以及
所述网络设备为所述终端设备配置上行资源，使得所述终端设备发送所述指示序列和所述CSI生成部输出比特序列。
48.根据附记39至47任一项所述的方法，其中，所述AI/ML模型为双边模型，具有模型标识和版本标识；所述一个双边模型的CSI生成部和所述CSI重构部具有同样的模型标识和版本标识，所述CSI生成部和所述CSI重构部具有不同的子标识。
49.根据附记39至48任一项所述的方法，其中，所述方法还包括：
所述网络设备向所述终端设备发送AI/ML相关能力查询；以及
所述网络设备接收所述终端设备反馈的AI/ML相关能力响应。
50.根据附记49所述的方法，其中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML训练能力信息、AI/ML升级能力信息。
51.根据附记49所述的方法，其中，所述方法还包括：
在所述网络设备确定所述终端设备具有AI/ML能力但所述CSI生成部的模型标 识和/或版本信息与所述CSI重构部的模型标识和/或版本信息不同时，指示或配置所述终端设备发送与所述CSI生成部的模型标识和/或版本信息对应的训练用CSI比特序列和/或对应的码书向量。
52.根据附记51所述的方法，其中，所述CSI的格式和/或大小被预定义，或者，所述CSI的格式和/或大小由所述网络设备配置。
53.根据附记51所述的方法，其中，所述方法还包括：
在所述CSI重构部利用所述CSI比特信息和所述码书向量完成训练后，所述网络设备将所述CSI重构部的模型标识和/或版本信息设置为与所述CSI生成部的模型标识和/或版本信息相同。
54.根据附记53所述的方法，其中，所述方法还包括：
所述网络设备向所述终端设备发送使能所述CSI生成部和/或所述CSI重构部的指示信息，和/或，发送确认所述CSI生成部可以使用的确认信息。
55.根据附记39至54任一项所述的方法，其中，所述方法还包括：
所述网络设备对所述码书向量进行扩充，并将所述扩充后的码书向量作为所述标签数据。
56.根据附记55所述的方法，其中，所述扩充后的码书向量中，W1的可选波束数目为最多6、8、10、12、16；W2的映射角度为最多5比特，幅度为最多5比特。
57.一种数据获取方法，其中，AI/ML模型包括位于终端设备内的CSI生成部和位于网络设备内的CSI重构部，所述方法包括：
所述终端设备获取所述CSI重构部输出的码书向量；以及
所述终端设备获取被输入所述CSI重构部的、并且与所述码书向量对应的CSI比特信息。
58.根据附记57所述的方法，其中，所述方法还包括：
所述终端设备将所述码书向量输入到所述CSI生成部中，并将所述CSI比特信息作为标签数据对所述CSI生成部进行训练。
59.根据附记57或58所述的方法，其中，所述CSI比特信息(Y)为输入到所述CSI重构部且对应所述码书向量的信息；所述码书向量(X)为特定数据集的部分或全部数据，由所述CSI比特信息(Y)被输入到所述CSI重构部后生成。
60.根据附记57至59任一项所述的方法，其中，所述CSI比特信息由控制信道 或数据信道承载并经由空口传输，和/或，所述码书向量根据预定义规则通过数据索引或指示序列或预编码矩阵指示(PMI)生成。
61.根据附记57至59任一项所述的方法，其中，
所述终端设备根据与所述CSI重构部相关的标识信息从内存或从所述终端设备内部或从所述终端设备外部获得所述CSI比特信息和/或所述码书向量。
62.根据附记61所述的方法，其中，所述与CSI重构部相关的标识信息包括：所述CSI重构部对应的模型标识和/或版本信息，和/或，所述重构生成部对应的模型的数据配置信息，和/或，所述CSI重构部的数据索引，和/或，所述CSI重构部的码书向量或DFT向量的指示序列。
63.根据附记57至62任一项所述的方法，其中，所述方法还包括：
所述终端设备接收所述网络设备发送的针对模型标识的模型训练指示和配置信息；所述配置信息包括训练用码书向量或DFT矩阵向量的指示序列以及所述向量对应的CSI重构部输入比特序列。
64.根据附记63所述的方法，其中，所述方法还包括：
所述终端设备基于所述指示序列，按照预定义规则生成相应的码书向量或DFT矩阵向量。
65.根据附记63或64所述的方法，其中，所述方法还包括：
所述终端设备接收所述网络设备的资源配置；以及
所述终端设备根据所述资源配置接收所述指示序列和所述CSI重构部输入比特序列。
66.根据附记57至65任一项所述的方法，其中，所述AI/ML模型为双边模型，具有模型标识和版本标识；同一AI/ML模型的CSI生成部和CSI重构部具有同样的模型标识和版本标识，所述CSI生成部和所述CSI重构部具有不同的子标识。
67.根据附记57至66任一项所述的方法，其中，所述方法还包括：
所述终端设备接收所述网络设备发送的AI/ML相关能力查询；以及
所述终端设备向所述网络设备反馈AI/ML相关能力响应。
68.根据附记67所述的方法，其中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML训练能力信息、AI/ML升级能力信息。
69.根据附记67所述的方法，其中，在所述网络设备确定所述终端设备具有AI/ML能力但所述CSI生成部的模型标识和/或版本信息与所述CSI重构部的模型标识和/或版本信息不同时，指示或配置所述终端设备接收与所述CSI重构部的模型标识和/或版本信息对应的CSI比特信息和/或码书向量。
70.根据附记69所述的方法，其中，所述CSI的格式和/或大小被预定义，或者，所述CSI的格式和/或大小由所述网络设备配置。
71.根据附记70所述的方法，其中，所述方法还包括：
在所述CSI生成部利用所述码书向量和所述CSI比特信息完成训练后，所述终端设备将所述CSI生成部的模型标识和/或版本信息设置为与所述CSI重构部的模型标识和/或版本信息相同。
72.根据附记71所述的方法，其中，所述方法还包括：
所述终端设备向所述网络设备发送确认其CSI生成部可以使用的确认信息,和/或，指示所述CSI生成部更新完成的指示信息。
73.根据附记57至72任一项所述的方法，其中，所述方法还包括：
所述终端设备对所述码书向量进行扩充，并将所述扩充后的码书向量作为所述标签数据。
74.根据附记73所述的方法，其中，所述扩充后的码书向量中，W1的可选波束数目为最多6、8、10、12、16；W2的映射角度为最多5比特，幅度为最多5比特。
75.一种网络设备，包括存储器和处理器，所述存储器存储有计算机程序，所述处理器被配置为执行所述计算机程序而实现如附记1至56任一项所述的数据获取方法。
76.一种终端设备，包括存储器和处理器，所述存储器存储有计算机程序，所述处理器被配置为执行所述计算机程序而实现如附记1至38、57-74任一项所述的数据获取方法。

Claims (20)

  1. 一种数据获取装置，配置于第二设备；其中，AI/ML模型包括位于第一设备内的信息生成部和位于所述第二设备内的信息重构部，所述数据获取装置包括：
    第一获取单元，其获取被输入所述信息生成部的第一数据；以及
    第二获取单元，其获取与所述第一数据对应并从所述信息生成部输出的第二数据。
  2. 根据权利要求1所述的装置，其中，所述装置还包括：
    第一训练单元，其将所述第二数据输入到所述信息重构部中，并将所述第一数据作为标签数据对所述信息重构部进行训练。
  3. 根据权利要求1所述的装置，其中，所述第一数据为输入到所述信息生成部的特定数据集的部分或全部数据；所述第二数据为将所述第一数据输入到所述信息生成部后生成的对应数据。
  4. 根据权利要求1所述的装置，其中，所述第一设备内的所述信息生成部和所述第二设备内的所述信息重构部使用所述第一数据和所述第二数据分别进行训练；
    所述第一设备内已训练的信息生成部在输入所述第一数据时，输出为所述第二数据或为近似所述第二数据的数据；所述第二设备内已训练的信息重构部在输入所述第二数据时，输出为所述第一数据或为近似所述第一数据的数据。
  5. 根据权利要求1所述的装置，其中，所述第一数据和所述第二数据是成对的数据集；所述成对的数据集具有模型标识信息。
  6. 根据权利要求4所述的装置，其中，一个第一数据对应多个第二数据。
  7. 根据权利要求1所述的装置，其中，所述第一获取单元或所述第二获取单元根据与所述信息生成部相关的标识信息从所述第二设备内部或从所述第二设备外部获得所述第一数据和/或所述第二数据。
  8. 根据权利要求7所述的装置，其中，与所述信息生成部相关的标识信息包括：所述信息生成部对应的模型标识和/或版本信息，和/或，所述信息生成部对应的模型的数据配置信息。
  9. 根据权利要求1所述的装置，其中，所述第二数据通过控制信道或数据信道承载并经由空口传输，或者，所述第二数据根据预定义规则通过数据索引生成，或者第二数据的索引由所述第一设备通过空口发送给所述第二设备。
  10. 根据权利要求1所述的装置，其中，所述第一数据被预先存储在所述第二设 备内，或者，所述第一数据根据预定义规则通过数据索引生成，或者，所述第一数据由所述第一设备通过空口发送给所述第二设备，或者第一数据的索引由所述第一设备通过空口发送给所述第二设备。
  11. 根据权利要求1所述的装置，其中，所述AI/ML模型具有模型标识和版本标识；同一AI/ML模型的信息生成部和信息重构部使用相同的模型标识和版本标识，所述信息生成部和所述信息重构部具有不同的子标识。
  12. 根据权利要求1所述的装置，其中，所述装置还包括：
    发送单元，其向所述第一设备发送AI/ML相关能力查询；以及
    接收单元，其接收所述第一设备反馈的AI/ML相关能力响应。
  13. 根据权利要求12所述的装置，其中，所述AI/ML相关能力包括如下至少之一：信号处理模块信息、AI/ML支持信息、AI/ML模型标识信息、版本信息、数据配置信息、AI/ML支持训练能力信息、AI/ML升级能力信息。
  14. 根据权利要求12所述的装置，其中，
    在所述第二设备确定所述第一设备具有AI/ML能力但所述信息生成部的模型标识和/或版本信息与所述信息重构部的模型标识和/或版本信息不同时，所述第一获取单元或所述第二获取单元获取与所述第一设备的信息生成部的模型标识和/或版本信息对应的所述第一数据和/或所述第二数据。
  15. 根据权利要求14所述的装置，其中，
    所述信息重构部在利用所述第一数据和所述第二数据完成训练后，所述信息重构部的模型标识和/或版本信息被设置为与所述信息生成部的模型标识和/或版本信息相同。
  16. 根据权利要求15所述的装置，其中，
    所述发送单元还向所述第一设备发送确认其信息生成部可以使用的确认信息,和/或，发送使能所述信息生成部和/或所述信息重构部的指示信息。
  17. 一种数据获取装置，配置于第一设备；其中，AI/ML模型包括位于所述第一设备内的信息生成部和位于第二设备内的信息重构部，所述装置包括：
    第三获取单元，其获取所述信息重构部输出的第一数据；以及
    第四获取单元，其获取被输入到所述信息重构部的、并且与所述第一数据对应的第二数据。
  18. 根据权利要求17所述的装置，其中，所述装置还包括：
    第二训练单元，其将所述第一数据输入到所述信息生成部中，并将所述第二数据作为标签数据对所述信息生成部进行训练。
  19. 根据权利要求17所述的装置，其中，所述第二数据为输入到所述信息重构部且对应所述第一数据的数据；所述第一数据为特定数据集的部分或全部数据，由所述第二数据被输入到所述信息重构部后生成。
  20. 一种通信系统，其中，AI/ML模型包括位于终端设备内的信道状态信息生成部和位于网络设备内的信道状态信息重构部，所述系统包括：
    网络设备，其获取被输入所述信道状态信息生成部的特定的码书向量；以及获取与所述码书向量对应的从所述信道状态信息生成部输出的信道状态信息比特信息；
    终端设备，其获取所述信道状态信息重构部输出的码书向量；以及获取被输入所述信道状态信息重构部的、并且与所述码书向量对应的信道状态信息比特信息。
PCT/CN2022/112299 2022-08-12 2022-08-12 数据获取方法与装置 Ceased WO2024031701A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
CN202280098380.XA CN119586082A (zh) 2022-08-12 2022-08-12 数据获取方法与装置
PCT/CN2022/112299 WO2024031701A1 (zh) 2022-08-12 2022-08-12 数据获取方法与装置
EP22954653.6A EP4572243A4 (en) 2022-08-12 2022-08-12 METHOD AND APPARATUS FOR DATA ACQUISITION
US19/029,817 US20250192849A1 (en) 2022-08-12 2025-01-17 Method and apparatus for acquiring data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/112299 WO2024031701A1 (zh) 2022-08-12 2022-08-12 数据获取方法与装置

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US19/029,817 Continuation US20250192849A1 (en) 2022-08-12 2025-01-17 Method and apparatus for acquiring data

Publications (1)

Publication Number Publication Date
WO2024031701A1 true WO2024031701A1 (zh) 2024-02-15

Family

ID=89850316

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/112299 Ceased WO2024031701A1 (zh) 2022-08-12 2022-08-12 数据获取方法与装置

Country Status (4)

Country Link
US (1) US20250192849A1 (zh)
EP (1) EP4572243A4 (zh)
CN (1) CN119586082A (zh)
WO (1) WO2024031701A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2026032812A1 (en) * 2024-08-07 2026-02-12 Aumovio Germany Gmbh Method of segmentation of model identification

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20240154709A1 (en) * 2022-11-03 2024-05-09 Apple Inc. Method of life cycle management using model id and model function

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20220041031A (ko) * 2020-09-24 2022-03-31 한국전자통신연구원 단말과 엣지 클라우드 서버 간의 분산 훈련 방법
CN114679355A (zh) * 2020-12-24 2022-06-28 华为技术有限公司 通信方法和装置
CN114697984A (zh) * 2020-12-28 2022-07-01 中国移动通信有限公司研究院 信息传输方法、终端及网络设备
CN114726413A (zh) * 2021-01-04 2022-07-08 中国移动通信有限公司研究院 信道信息获取方法、装置及相关设备

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20220041031A (ko) * 2020-09-24 2022-03-31 한국전자통신연구원 단말과 엣지 클라우드 서버 간의 분산 훈련 방법
CN114679355A (zh) * 2020-12-24 2022-06-28 华为技术有限公司 通信方法和装置
CN114697984A (zh) * 2020-12-28 2022-07-01 中国移动通信有限公司研究院 信息传输方法、终端及网络设备
CN114726413A (zh) * 2021-01-04 2022-07-08 中国移动通信有限公司研究院 信道信息获取方法、装置及相关设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4572243A4 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2026032812A1 (en) * 2024-08-07 2026-02-12 Aumovio Germany Gmbh Method of segmentation of model identification

Also Published As

Publication number Publication date
EP4572243A4 (en) 2025-11-26
EP4572243A1 (en) 2025-06-18
CN119586082A (zh) 2025-03-07
US20250192849A1 (en) 2025-06-12

Similar Documents

Publication Publication Date Title
CN114747250B (zh) 信道状态信息的反馈
CN114762276A (zh) 信道状态信息反馈
CN114616762A (zh) 用于传输信道状态信息的方法和设备
US20250192849A1 (en) Method and apparatus for acquiring data
CN117203922A (zh) 用于多trp csi-rs报告的cmr和imr配置增强
US20240372751A1 (en) Channel estimation method, method for training channel estimation model, and communication device
WO2024067258A1 (zh) 一种通信的方法和通信装置
WO2021155609A1 (zh) 信号发送方法及装置
WO2023115254A1 (zh) 处理数据的方法及装置
WO2024169757A1 (zh) 一种通信的方法和通信装置
TW202315347A (zh) 用於 mimo csi 回饋的方法和裝置
WO2023236143A1 (zh) 信息收发方法与装置
WO2025139843A1 (zh) 通信方法和通信装置
WO2022151064A1 (zh) 信息发送方法、信息接收方法、装置、设备及介质
US20240048207A1 (en) Method and apparatus for transmitting and receiving feedback information based on artificial neural network
WO2024208296A1 (zh) 一种通信的方法和通信装置
WO2025039261A1 (zh) 一种数据处理方法及装置
WO2024032775A1 (zh) 量化方法和装置
WO2025035433A1 (zh) 一种通信方法及通信装置
CN117678257A (zh) 信息处理方法、模型生成方法及设备
US20250053874A1 (en) Method and apparatus for sequential learning of two-sided artificial intelligence/machine learning model for feedback of channel state information in communication system
CN121794930A (zh) 用于双侧模型配对的方法和装置
US20250055526A1 (en) Givens rotation matrix parameterization pre-processing for channel state information feedback enhancement in a communication network
WO2025231790A1 (en) Mechanisms for model identification and offline engineering
WO2026065514A1 (zh) Ai/ml模型或功能的配置方法和装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22954653

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 202280098380.X

Country of ref document: CN

WWP Wipo information: published in national office

Ref document number: 202280098380.X

Country of ref document: CN

WWE Wipo information: entry into national phase

Ref document number: 2022954653

Country of ref document: EP

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2022954653

Country of ref document: EP

Effective date: 20250312

WWP Wipo information: published in national office

Ref document number: 2022954653

Country of ref document: EP