WO2024021621A1 - 功率放大器模型的获取方法、装置及功率放大器模型 - Google Patents

功率放大器模型的获取方法、装置及功率放大器模型 Download PDF

Info

Publication number
WO2024021621A1
WO2024021621A1 PCT/CN2023/080821 CN2023080821W WO2024021621A1 WO 2024021621 A1 WO2024021621 A1 WO 2024021621A1 CN 2023080821 W CN2023080821 W CN 2023080821W WO 2024021621 A1 WO2024021621 A1 WO 2024021621A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
power amplifier
input data
sub
obtaining
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/CN2023/080821
Other languages
English (en)
French (fr)
Inventor
杨振
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.)
ZTE Corp
Original Assignee
ZTE Corp
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 ZTE Corp filed Critical ZTE Corp
Priority to US18/875,106 priority Critical patent/US20250315727A1/en
Priority to KR1020257000396A priority patent/KR20250018557A/ko
Priority to EP23844839.3A priority patent/EP4546204A4/en
Publication of WO2024021621A1 publication Critical patent/WO2024021621A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/30—Circuit design
    • G06F30/36—Circuit design at the analogue level
    • G06F30/367—Design verification, e.g. using simulation, simulation program with integrated circuit emphasis [SPICE], direct methods or relaxation methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • 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
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0499—Feedforward 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
    • H—ELECTRICITY
    • H03—ELECTRONIC CIRCUITRY
    • H03F—AMPLIFIERS
    • H03F3/00—Amplifiers with only discharge tubes or only semiconductor devices as amplifying elements
    • H03F3/20—Power amplifiers, e.g. Class B amplifiers, Class C amplifiers
    • H03F3/21—Power amplifiers, e.g. Class B amplifiers, Class C amplifiers with semiconductor devices only
    • H—ELECTRICITY
    • H03—ELECTRONIC CIRCUITRY
    • H03F—AMPLIFIERS
    • H03F3/00—Amplifiers with only discharge tubes or only semiconductor devices as amplifying elements
    • H03F3/20—Power amplifiers, e.g. Class B amplifiers, Class C amplifiers
    • H03F3/21—Power amplifiers, e.g. Class B amplifiers, Class C amplifiers with semiconductor devices only
    • H03F3/213—Power amplifiers, e.g. Class B amplifiers, Class C amplifiers with semiconductor devices only in integrated circuits

Definitions

  • Embodiments of the present application relate to the field of signal processing technology, and in particular, to a method and device for obtaining a power amplifier model, a power amplifier model, a storage medium, and a program product.
  • Embodiments of the present application provide a method and device for obtaining a power amplifier model, a power amplifier model, a storage medium, and a program product, aiming to improve the generalization ability of the power amplifier model based on neural networks.
  • embodiments of the present application provide a method for obtaining a power amplifier model.
  • the method includes: obtaining an initial sub-model of a power amplifier, label data, and input data; based on the label data and the input data, The initial sub-model is iteratively trained until the iteration stop condition is reached; and after each iterative training is completed, a target sub-model is obtained; a power amplifier model is obtained based on at least one of the target sub-models.
  • embodiments of the present application provide a device for obtaining a power amplifier model, including: a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, the following is implemented: The method for obtaining the power amplifier model described in the first aspect.
  • embodiments of the present application provide a power amplifier model, which is obtained according to the method for obtaining a power amplifier model described in the first aspect.
  • embodiments of the present application provide a computer-readable storage medium, including: the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to execute the power as described in the first aspect. How to obtain the amplifier model.
  • embodiments of the present application provide a computer program product, including a computer program or computer instructions.
  • the computer program or computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device can read the computer program from the computer. Reading the storage medium reads the computer program or the computer instructions, and the processor executes the computer program or the computer instructions, so that the computer device performs the method for obtaining the power amplifier model as described in the first aspect. Law.
  • multiple target sub-models are obtained by performing multiple iterative training on the initial sub-model, and the obtained multiple The model composed of the target sub-model is used as the final power amplifier model.
  • the model structure and training process By optimizing the model structure and training process, the generalization ability of the power amplifier model and the prediction accuracy of the model are improved.
  • Figure 1 is an application environment diagram of the power amplifier modeling method in an embodiment of the present application
  • Figure 2 is a schematic diagram of a power amplifier modeling provided by an embodiment of the present application.
  • Figure 3 is a flow chart of a method for obtaining a power amplifier model provided by an embodiment of the present application
  • Figure 4 is a flow chart for calculating the priority of input data at historical moments provided by an embodiment of the present application
  • Figure 5 is a flow chart for calculating the priority of input data at historical moments provided by another embodiment of the present application.
  • Figure 6 is a flow chart for obtaining a pre-trained neural network model provided by an embodiment of the present application.
  • Figure 7 is a flow chart for selecting a target sub-model to construct a power amplifier model provided by an embodiment of the present application
  • Figure 8 is a schematic structural diagram of a power amplifier model provided in an example of this application.
  • Figure 9 is a schematic structural diagram of a power amplifier model provided by another example of this application.
  • Figure 10 is a schematic structural diagram of a power amplifier model provided by another example of this application.
  • Figure 11 is a schematic model structure diagram of an initial sub-model provided in an example of this application.
  • Figure 12(a) is a model structure diagram of a unit in an example initial sub-model of this application.
  • Figure 12(b) is a model structure diagram of a unit in an example initial sub-model of this application.
  • Figure 12(c) is a model structure diagram of a unit in an example initial sub-model of this application.
  • Figure 12(d) is a model structure diagram of a unit in an example initial sub-model of this application.
  • Figure 13 is a structural diagram of a power amplifier model provided by an embodiment of the present application.
  • Figure 14 is a structural diagram of a power amplifier model provided by an embodiment of the present application.
  • FIG. 15 is a schematic structural diagram of a device for obtaining a power amplifier model provided by an embodiment of the present application.
  • the radio frequency power amplifier is an important component in the wireless communication system. Its function is to amplify the power of the wireless signal to a target value and then feed it into the antenna.
  • RF power amplifiers have two significant characteristics: nonlinearity and memory. Nonlinearity is mainly manifested in the fact that the gain value of the power amplifier to the input signal changes as the input signal power increases, rather than being constant. Nonlinearity will cause distortion of the signal constellation diagram, especially for signals with Peak to Average Power Ratio (PAPR) such as Orthogonal Frequency Division Multiplexing (OFDM).
  • PAPR Peak to Average Power Ratio
  • OFDM Orthogonal Frequency Division Multiplexing
  • Memory means that the output of the power amplifier is not only related to the input at the current moment, but also related to the input signal at the historical moment. The impact of memory is related to the bandwidth of the signal. The greater the bandwidth of the signal, the greater the impact of memory and the more it cannot be ignored.
  • Behavioral modeling of power amplifiers is the basis of power amplifier-related research work, such as digital pre-distortion (DPD).
  • Power amplifier modeling specifically simulates the nonlinearity and memory of the power amplifier to establish the mathematical relationship between the input and output of the power amplifier.
  • models used for modeling power amplifiers can be roughly divided into two categories.
  • One is traditional models, such as the MP model and GMP model based on the Volterra model; the other is models based on neural networks, such as MLP.
  • the traditional power amplifier model is designed with the characteristics of the power amplifier fully considered. For example, the high-order terms in the model correspond to the intermodulation components in the power amplifier product. Therefore, it matches the actual power amplifier very well and has good Generalization.
  • the distortion principle of power amplifiers has become complex, and the modeling accuracy of traditional models has dropped significantly.
  • researchers have begun to turn their attention to methods based on neural networks.
  • the neural network is a universal model, and power amplifier modeling is only one of its application scenarios. Since the design of the model does not consider the distortion characteristics of the power amplifier, models based on neural networks generally have the defect of poor generalization ability, that is, they have high fitting accuracy on the training set, but have a large fitting accuracy on the test set. Decay, especially when the number of training set samples is small.
  • embodiments of the present application provide a method and device for obtaining a power amplifier model, a power amplifier model, a storage medium and a program product.
  • a method and device for obtaining a power amplifier model, a power amplifier model, a storage medium and a program product By performing multiple iterative trainings on the initial sub-model, multiple target sub-models are obtained, and The model composed of multiple target sub-models is used as the final power amplifier model.
  • the model structure and training process By optimizing the model structure and training process, the generalization ability of the power amplifier model and the prediction accuracy of the model are improved.
  • the method for obtaining the power amplifier model provided by this application can be applied in the application environment described in Figure 1.
  • the electronic device 100 obtains the input and output sample values of the power amplifier 200, and then models the input and output characteristics of the power amplifier to obtain the power amplifier model 300.
  • the above-mentioned electronic device may be a predistortion module in a base station or a signal processing unit in a radar system.
  • the embodiments of the present application are not limited here.
  • the above-mentioned power amplifier may be a power amplifier component in a base station, a power amplifier unit in a radar system, or a power amplifier device in a terminal device, etc.
  • the embodiments of the present application are not limited here.
  • Figure 2 is a schematic diagram of a power amplifier modeling provided by an embodiment of the present application. As shown in the figure, assuming that the input and output signals of the power amplifier are x and y respectively, the principle of modeling it is shown in Figure 2. The goal is to make the output of the established power amplifier model The error e between the output y of PA and PA is the smallest.
  • FIG. 3 is a flow chart of a method for obtaining a power amplifier model provided by an embodiment of the present application. As shown in FIG. 3 , the method for obtaining the power amplifier model may include but is not limited to step S1000, step S2000, and step S3000.
  • Step S1000 Obtain the initial sub-model, label data and input data of the power amplifier.
  • the initial sub-model of the power amplifier is a neural network model
  • the label data is the power amplifier
  • the error between the actual output signal and the initial sub-model output signal of the power amplifier, the input data is the input signal at the current moment and the input signal at the historical moment. Since it is considered that the input signal at the historical moment has an impact on the output at the current moment, in the embodiment of the present application, when acquiring the model input data, the input signal at the historical moment is obtained and used as input data for model training. In the field of power amplifier technology, memory depth is often used to indicate how long ago the current output signal is related to the input signal.
  • the initial sub-model of the power amplifier is a model composed of multiple neural network models.
  • the initial sub-model composed of multiple neural network models may have more advantages in fitting the data.
  • the initial sub-model can be a neural network model or a model composed of multiple neural network models, and different initial sub-models can be selected according to different needs of application scenarios.
  • Step S2000 Iteratively train the initial sub-model according to the label data and input data until the iteration stop condition is reached; and after each iterative training is completed, a target sub-model is obtained.
  • the initial sub-model is iteratively trained using label data and input data.
  • this iterative training is stopped and a trained target sub-model is obtained.
  • the label data is updated according to the output of the target sub-model to obtain new label data.
  • the new label data is used for the next sub-model iterative training.
  • the iteration stop condition is a preset number of iterations. Therefore, when the current number of iterations is less than the preset number of iterations, input data at the current time and at least one historical time input data are input to the initial sub-model for execution. Continuous iterative training, and after each iterative training, a target sub-model is generated. Multiple target sub-models obtained through multiple iterations of training have the opportunity to constitute the final power amplifier model.
  • the iteration stop condition is a preset number of iterations. Therefore, when the current number of iterations is less than the preset number of iterations, the input data at the current time and all historical time input data are input to the initial sub-model for continuous processing. Iterative training, and after each iteration of training, a target sub-model is generated. Multiple target sub-models obtained through multiple iterations of training have the opportunity to constitute the final power amplifier model.
  • the iteration stop condition is a preset number of iterations. Therefore, when the current number of iterations is less than the preset number of iterations, during each iteration training process, the input data at the current moment and part of the historical moments are Input data is input to the initial sub-model, and until the preset number of iterations is reached, the iterative training is stopped, and a target sub-model is generated after each iterative training. Multiple target sub-models obtained through multiple iterations of training have the opportunity to constitute the final power amplifier model.
  • the iteration stop condition is that all historical moment input data participate in model iterative training. Therefore, in the current iterative training process, the priority of each historical moment input data is generated according to the preset priority calculation conditions, Based on the number of iterative training of the current sub-model and the priority of the input data at historical moments, a historical moment input data set is constructed. When the input data of the current iterative training does not reach the iteration stop condition, the current moment input data and the historical moment input data set are Input to the initial sub-model for iterative training.
  • the historical moment input data needs to be sorted according to the priority of the historical moment input data, to obtain the sorted historical moment input data, and then to input data at the sorted historical moment. , select the target number of historical moment input data in turn to construct a historical moment input data set.
  • the target number is equal to the number of iterative training of the current sub-model. It can be understood that as the number of iterative training of the sub-model increases, more and more The historical moment input data are selected to form the historical moment input data set, until all the historical moment data are selected to form the historical moment input data set, and it ends at this time Iterative training of submodels.
  • the iterative training process of the sub-model needs to be performed Q+1 times. , and finally generate Q+1 target sub-models.
  • the historical moment input data set for example, during each iteration of training, two or more top-ranked historical moment input data are selected and added to the previous iteration.
  • the historical moment input data set constructed by training is formed into a new historical moment input data set.
  • Figure 4 is a flow chart for calculating the priority of input data at historical moments provided by an embodiment of the present application.
  • calculating historical time input data includes at least steps S2110, S2120 and S2130.
  • Step S2110 Obtain the error value of the current sub-model iterative training.
  • the residual signal in the current iterative training process is calculated, and the residual signal is the difference between the actual output of the power amplifier and the predicted output of the model.
  • Step S2120 Calculate the correlation between the error value and the historical time input data.
  • the correlation between the error value and the historical moment input data is calculated according to the correlation calculation formula, and the correlation value is obtained.
  • any method that can calculate the correlation between the error value and the input data at historical moments can be used in the correlation calculation in step S2120 in this embodiment of the present application.
  • the embodiments of the present application are not limited here.
  • Step S2130 Obtain the priority of the input data at historical moments based on the correlation.
  • the obtained correlation values are sorted.
  • the level will be higher. Prioritizing historical moment input data with high priority to construct a historical moment input data set can improve the generation efficiency and prediction accuracy of the final power amplifier model.
  • the embodiment of Figure 4 sequentially constructs multiple historical time input data sets with different historical time input data according to the priority of historical time input data, uses the historical time input data sets to perform iterative training of sub-models, and generates multiple target sub-models. , which can effectively improve the training efficiency of the model and the generalization ability of the model in practical application scenarios.
  • Figure 5 is a flow chart for calculating the priority of input data at historical moments provided by another embodiment of the present application.
  • calculating historical time input data includes at least steps S2210, S2220 and S2230.
  • Step S2210 Obtain the pretrained neural network model.
  • a pre-trained neural network model is obtained as a prediction model for fitting accuracy of input data at historical moments.
  • the pre-trained neural network model is a model trained with training data, which can generate fitting accuracy corresponding to the input data at the input historical moment during actual prediction.
  • Figure 6 is a flow chart for obtaining a pre-trained neural network model provided by an embodiment of the present application.
  • Get a pretrained neural network The network model specifically includes steps S2211 and S2212.
  • Step S2211 Construct multiple training data sets through combination according to the input data at the current time and the input data at at least one historical time.
  • the input data in the training data used is a combination of input data at the current moment and input data at historical moments.
  • the label data of the training data is the error value between the actual output and the predicted output.
  • the priority of the input data at the historical moment is sorted, and the top-ranked input data at the historical moment is selected to construct the input data together with the input data at the current moment.
  • the constructed input data is called a temporary queue.
  • the temporary queue when building the first temporary queue, contains the input data at the current time and the input data at the historical time with the highest priority; when building the second temporary queue, the temporary queue contains the input at the current time.
  • the data and the top two historical moment input data with priority, and so on, until all historical moment input data are selected to build a temporary queue.
  • this step is pre-training of the model, so you can build as many temporary queues as possible to obtain more training data, so that during the training process of the model, a model with more accurate prediction results can be generated.
  • part of the input data at historical moments can be selected to form training data with the current input data. These selected input data at historical moments can be randomly selected. of.
  • Step S2212 Input multiple training data sets to the neural network model respectively, and train the neural network model through the label data to obtain a pre-trained neural network model.
  • the model after selecting a neural network model, the model can be trained.
  • the training steps are the same as those of common neural networks, and will not be described in detail here.
  • the selection of the neural network model in this step can be adjusted according to actual needs, and the training of the neural network model can also use common supervised learning methods.
  • the embodiments of the present application are not limited here.
  • Step S2220 Input the historical moment input data to the pre-trained neural network model for fitting, and obtain a fitting accuracy value corresponding to the historical moment input data.
  • inputting historical time input data into a trained neural network model can predict the fitting accuracy value of the historical time input data.
  • Step S2230 Obtain the priority of input data at historical moments based on the fitting accuracy value.
  • the higher the fitting accuracy value the higher the priority corresponding to the input data at the historical time.
  • the lower the fitting accuracy value the lower the priority corresponding to the input data at the historical time.
  • Step S3000 Obtain a power amplifier model based on at least one target sub-model.
  • all target sub-models generated by sub-model iterative training are used to construct the final power amplifier model.
  • the power amplifier model obtained in this way includes the impact of input data at all historical moments on the model output, so it may have higher prediction accuracy in some application scenarios.
  • the network submodel can be set by setting thresholds on one or more statistical dimensions during the iterative training process of the submodel, such as mean squared error (MSE) or normalized mean squared error (NMSE). If the statistic is worse than this threshold after iterative training to convergence, it can be considered that this target sub-model is not helpful in building the final power amplifier model, so one or more target sub-models that do not meet the threshold requirements are discarded.
  • MSE mean squared error
  • NMSE normalized mean squared error
  • Figure 7 is a flow chart for selecting a target sub-model to construct a power amplifier model provided by an embodiment of the present application. As shown in the figure, it includes at least steps S3100, S3200, S3300 and S3400.
  • Step S3100 Obtain a preset statistical dimension and a preset statistical dimension threshold corresponding to the preset statistical dimension.
  • the mean square error is selected as the statistical dimension, and the mean square error threshold is preset in advance. If the mean square error of the target sub-model exceeds the preset mean square error threshold, it means that the training effect of the sub-model cannot meet expectations, and the target sub-model may be abandoned in the future, and other mean square errors will be selected that do not exceed the preset mean square error.
  • the target sub-model of the threshold is used to construct the final power amplifier model.
  • Step S3200 Obtain the statistical value corresponding to the preset statistical dimension of each target sub-model.
  • the mean square error of each target sub-model is obtained.
  • Step S3300 Determine the target sub-model whose statistical value is better than the preset statistical dimension threshold as the power amplifier sub-model.
  • target sub-models whose mean square error is less than or equal to the mean square error threshold are selected, and these target sub-models that meet the requirements are finally used to construct the power amplifier model.
  • the mean square error of all target sub-models is less than or equal to the mean square error threshold, in order to achieve the purpose of simplifying the model, a preset number of target sub-models with relatively smaller mean square error can be selected for processing. Construction of power amplifier model.
  • a partial target sub-model is also used to construct the final power amplifier model.
  • the model is not pre-set a threshold for the statistical dimension. screening, but directly pre-set the complexity of the final power amplifier model, and sort the convergence degree of the target sub-models according to the set complexity, discard one or more target sub-models, and make the remaining target sub-models
  • the power amplifier model composed of the model meets the preset model complexity.
  • Step S3400 Determine the target sub-model whose statistical value is better than the preset statistical dimension threshold as the power amplifier sub-model.
  • the power amplifier sub-models are sorted according to the mean square error statistical value to obtain the sorting order, and the model composed of the power amplifier sub-models according to the sorting order is determined as the power amplifier model.
  • embodiments that only use part of the target sub-models to construct the final power amplifier model can reduce the parameters of the model and reduce the complexity of the model by reducing the number of sub-models used in the model, thereby improving the generalization ability of the model.
  • Step S101 Obtain the initial sub-model, that is, initialize the sub-model NN iter .
  • Step S102 Generate input data X iter and label data Y iter of the initial sub-model NN iter , where the input data X iter and label data Y iter are generated as follows:
  • the current moment input data x(n) is used for training; in the second iteration process, the current moment input data x(n) and the historical moment input data x(n) with a memory depth of 1 are used. n-1) for training; in the third iteration process, the current moment input data x(n), the historical moment input data x(n-1) with a memory depth of 1, and the historical moment input data x with a memory depth of 2 are used (n-2) for training, and so on.
  • Step S103 Train the initial sub-model to convergence.
  • Step S104 Update the output of the power amplifier model according to the generated target sub-model.
  • Example 2 The difference between Example 2 and Example 1 is that the same initial sub-model is used. The difference is that the sub-model training method in Example 2 is different, as follows:
  • Step S201 Obtain the initial sub-model, that is, initialize the sub-model NN iter .
  • Step S202 Generate the input data X and label data Y of the sub-model neural network NN iter .
  • the generation method of X and Y is as follows;
  • Step S203 Train the sub-model neural network NN iter until convergence.
  • Step S204 Update the output of the power amplifier model according to the generated target sub-model.
  • Step S205 Determine whether queue S2 is empty. If it is empty, the process ends.
  • queue S 2 In the first iteration of training, only the input data at the current moment is used for model training, so queue S 2 must not be empty.
  • Step S206 Calculate the priority order of elements in queue S2 .
  • the correlation measurement method is used to calculate the priority of input data at historical moments in queue S2 .
  • the specific calculation process is as follows:
  • Sub-step 1 Calculate the residual signal r of the current iteration
  • Sub-step 2 Calculate the i-th element in queue S 2 related measures
  • Sub-step 3 Prioritize according to the correlation metric. The larger the correlation metric, the higher the priority.
  • Step S207 Add the element with the highest priority in queue S2 to the end of queue S1 , and delete the element from queue S2 .
  • one historical moment input data is added to the current moment input data each time to form the input data for this iterative training.
  • the memory depth is too large, it can be added each time Add two or more historical moment input data to the current moment input data, thereby saving model training time and improving model training efficiency.
  • the iterative training is stopped after Q+1 times.
  • the Q+1 target sub-models of NN 1 , NN 2 ,..., NN Q+1 are obtained. These Q+1 target sub-models together constitute the example of this application.
  • the proposed power amplifier model, the schematic diagram of the model is shown in Figure 9.
  • the priority sorting module in the figure is to sort and output x(n), x(n-1),...,x(nQ) according to the order of elements in queue S 1 , that is
  • Example 3 The similarity between Example 3 and Example 2 lies in the use of the same initial sub-model and training ideas.
  • the difference lies in the priority calculation method of input data at historical moments in Example 3. The details are as follows:
  • Step S301 Obtain the initial sub-model, that is, initialize the sub-model NN iter .
  • Step S302 Generate the input data X and label data Y of the sub-model neural network NN iter .
  • the generation method of X and Y is as follows;
  • Step S303 Train the sub-model neural network NN iter until convergence.
  • Step S304 Update the output of the power amplifier model according to the generated target sub-model.
  • Step S305 Determine whether queue S2 is empty. If it is empty, the process ends.
  • queue S 2 In the first iteration of training, only the input data at the current moment is used for model training, so queue S 2 must not be empty.
  • Step S306 Calculate the priority order of elements in queue S2 .
  • the correlation measurement method is used to calculate the priority of input data at historical moments in queue S2 .
  • the specific calculation process is as follows:
  • Sub-step 1 Generate temporary queue ⁇ , Represents the i-th element in queue S 2 ;
  • Sub-step 2 Initialize the neural network NN tmp .
  • the neural network NN tmp can be any type of neural network that meets the requirements.
  • Sub-step 3 Generate the input data X and label data Y of the neural network NN tmp .
  • X and Y are generated as follows:
  • ⁇ i represents the i-th element in queue ⁇ .
  • the purpose of generating the temporary queue is to train the neural network NN tmp so that the trained model NN tmp can accurately predict the fitting accuracy of the input data at historical moments during the application process, thereby obtaining which historical The input data at all times has a greater impact on the output results, thus giving it a higher priority.
  • Sub-step 4 Use the above input data and label data to train the neural network NN tmp .
  • Sub-step 5 Calculate the fitting accuracy of NN tmp
  • Sub-step 6 Prioritize according to the fitting accuracy. The higher the fitting accuracy, the higher the priority.
  • Step S307 Add the element with the highest priority in queue S2 to the end of queue S1 , and delete the element from queue S2 .
  • one historical moment input data is added to the current moment input data each time to form the input data for this iterative training.
  • the memory depth is too large, it can be added each time Add two or more historical moment input data to the current moment input data, thereby saving model training time and improving High model training efficiency.
  • the iterative training is stopped after Q+1 times.
  • the Q+1 target sub-models of NN 1 , NN 2 ,..., NN Q+1 are obtained. These Q+1 target sub-models together constitute the example of this application.
  • the proposed power amplifier model, the schematic diagram of the model is shown in Figure 9.
  • the priority sorting module in the figure is to sort and output x(n), x(n-1),..., x(nQ) according to the order of elements in queue S 1 , that is
  • the complexity of the model needs to be simplified to make it suitable for more application scenarios. Therefore, all of the multiple target sub-models obtained in the foregoing Examples 1, 2, and 3 may not be selected to construct the final power amplifier model. In this case, these target sub-models need to be selected.
  • the specific method is to set a threshold for a certain statistical dimension in the neural network NN iter training.
  • This statistical dimension can be the mean square error MSE or the standard mean square error NMSE. If the statistics of the network NN iter are worse than this threshold after training to convergence, it can be considered that NN iter is not helpful in constructing Y, so NN iter is unnecessary, and NN iter does not participate in the construction of the final power amplifier model.
  • Example 4 can still use the initial sub-model selected in Example 1, but the difference is that the target sub-model is screened during the training process, as follows:
  • Step S401 Initialize the neural network NN iter .
  • Step S402 Generate the input data X and label data Y of the neural network NN iter .
  • the generation method of X and Y is as follows:
  • Step S403 Train the neural network NN iter until convergence.
  • Step S404 Determine the statistics and threshold of NN iter relationship, if it is better than the set threshold, the NN iter is deemed to be valid, and the output of the power amplifier model is updated. If it is worse than the set threshold, the NN iter is deemed invalid, and the NN iter does not participate in the construction of the final power amplifier model.
  • Step S405 Determine whether queue S2 is empty. If it is empty, the process ends.
  • Step S406 Calculate the priority order of elements in queue S2 .
  • Step S407 Add the element with the highest priority in queue S2 to the end of queue S1 , and delete the element from queue S2 .
  • the memory depth Q is set to 4, and the target statistic is the standard mean square error NMSE.
  • the calculation formula of standard mean square error NMSE is standard mean square error
  • the elements of queue S 1 are ⁇ 0,1,4,2,3 ⁇ .
  • the statistical index standard mean square error NMSE of NN 2 and NN 4 is higher than the set threshold, then NN 2 and NN are considered 4 is invalid and does not participate in the construction of the final power amplifier model.
  • the final power amplifier model is shown in Figure 10.
  • this example can reduce the number of target sub-models used by the model, thereby reducing the parameters of the model and reducing the complexity of the model.
  • Figure 11 is a schematic diagram of the model structure of the initial sub-model provided in the example of this application.
  • the main structure of unit 1 is a complex multilayer perceptron (MLP) network, and the absolute value of the output and input of unit 1 constitutes the input of the MLP network.
  • Unit 2 multiplies the output of unit 1 and X and then sums them.
  • Unit 3 performs a weighted summation of the output of unit 1, where the weight coefficients w 1 , w 2 ,..., w iter are parameters of the network and need to be trained.
  • Unit 4 is a linear model, which is essentially a weighted summation of X, in which the weight coefficients ⁇ 1 , ⁇ 2 ,..., ⁇ iter are parameters of the network and need to be trained.
  • Figures 12(a)-(d) are provided, which are model structure diagrams of each unit in the initial sub-model of the example of this application.
  • initial sub-model structure provided in this example is only an example, and those skilled in the art can select an appropriate initial sub-model structure for training according to actual needs.
  • the method for obtaining the power amplifier model provided in this application example improves the generalization ability of the power amplifier model and the prediction accuracy of the model by optimizing the model structure and training process.
  • the final power amplifier model is obtained according to the method for obtaining a power amplifier model provided by the embodiment of the present application.
  • the verification process is as follows:
  • Step S501 Collect a set of input and output data x and y from the power amplifier.
  • Step S502 Extract the power amplifier model from this set of data using the traditional neural network-based power amplifier modeling method and the power amplifier modeling method obtained in the embodiment of the present application respectively.
  • Step S503 Calculate the power amplifier model obtained in step S502 to obtain a digital predistortion DPD model.
  • Step S504 Pass x through the DPD model obtained in step S503 to obtain the signal z, and then pass the signal z through the power amplifier to obtain the signal
  • Step S505 Through x and Calculate DPD performance statistics where G is the gain of the power amplifier.
  • the traditional neural network-based power amplifier modeling method is used to test signals with two bandwidths of 20M and 40M respectively.
  • the final statistics are shown in Table 1 below.
  • the neural network after the fourth iteration of model training cannot meet the threshold requirements, which is not helpful for building a power amplifier model.
  • the queue S 1 ⁇ 0,5,9,1,3,6,2,8,4,7 ⁇ , and the final power amplifier model is shown in Figure 13.
  • the neural network after the fifth iteration of model training cannot meet the threshold requirements, which is not helpful for building a power amplifier model.
  • the queue S 1 ⁇ 0,2,6,9,8,4,5,7,3 ⁇ , and the final power amplifier model is shown in Figure 14.
  • FIG. 15 is a schematic structural diagram of a device for obtaining a power amplifier model provided by an embodiment of the present application.
  • the device includes a memory 1100 and a processor 1200.
  • the number of memories 1100 and processors 1200 can be one or more.
  • one memory 1100 and one processor 1200 are taken as an example.
  • the memory 1100 and processor 1200 in the device can be connected through a bus or other means.
  • Figure 15 Take the example of connecting via a bus.
  • the memory 1100 can be used to store software programs, computer-executable programs and modules, such as program instructions/modules corresponding to the resource determination method provided in any embodiment of the present application.
  • the processor 1200 implements the above method of obtaining the power amplifier model by running software programs, instructions and modules stored in the memory 1100 .
  • the memory 1100 may mainly include a program storage area and a data storage area, where the program storage area may store an operating system and at least one application program required for a function.
  • the memory 1100 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid-state storage device.
  • memory 1100 may further include memory located remotely relative to processor 1200, and these remote memories may be connected to the device through a network. Examples of the above-mentioned networks include but are not limited to the Internet, intranets, local area networks, mobile communication networks and combinations thereof.
  • An embodiment of the present application also provides a computer-readable storage medium that stores computer-executable instructions.
  • the computer-executable instructions are used to execute the method for obtaining a power amplifier model as provided in any embodiment of the present application.
  • An embodiment of the present application also provides a computer program product, which includes a computer program or computer instructions.
  • the computer program or computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer program from the computer-readable storage medium.
  • the processor executes the computer program or the computer instruction, so that the computer device executes the method for obtaining the power amplifier model provided by any embodiment of the present application.
  • the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may consist of several physical components. Components execute cooperatively. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit . Such software may be distributed on computer-readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media).
  • computer storage media includes volatile and nonvolatile media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. removable, removable and non-removable media.
  • Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disk (DVD) or other optical disk storage, magnetic cassettes, tapes, disk storage or other magnetic storage devices, or may Any other medium used to store the desired information and that can be accessed by a computer.
  • communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media .
  • a component may be, but is not limited to, a process, processor, object, executable file, thread of execution, program or computer running on a processor.
  • calculating Applications running on the device and computing devices can both be components.
  • One or more components can reside in a process or thread of execution, and the component can be localized on one computer or distributed between 2 or more computers. Additionally, these components can execute from various computer-readable media having various data structures stored thereon.
  • a component may, for example, be based on a signal having one or more data packets (eg, data from two components interacting with another component, such as a local system, a distributed system, or a network, such as the Internet, which interacts with other systems via signals) Communicate through local or remote processes.
  • data packets eg, data from two components interacting with another component, such as a local system, a distributed system, or a network, such as the Internet, which interacts with other systems via signals

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Hardware Design (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Geometry (AREA)
  • Microelectronics & Electronic Packaging (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Power Engineering (AREA)
  • Amplifiers (AREA)

Abstract

本申请实施例提供了一种功率放大器模型的获取方法、装置及功率放大器模型、存储介质及程序产品，通过对初始子模型进行多次迭代训练，得到多个目标子模型，并将得到的多个目标子模型构成的模型作为最终的功率放大器模型。

Description

功率放大器模型的获取方法、装置及功率放大器模型
相关申请的交叉引用
本申请基于申请号为202210885045.6、申请日为2022年7月26日的中国专利申请提出，并要求该中国专利申请的优先权，该中国专利申请的全部内容在此引入本申请作为参考。
技术领域
本申请实施例涉及信号处理技术领域，尤其涉及一种功率放大器模型的获取方法、装置及功率放大器模型、存储介质及程序产品。
背景技术
近年来随着功率放大器技术的进步，新的种类的功率放大器失真原理变得更加复杂，传统模型受到模型本身的限制难以准确刻画功率放大器的失真特性。相关技术中，虽然提出了基于神经网络的功率放大器模型的相关概念，但是在实际应用中仍存在诸多问题，例如模型的泛化能力差，即在训练数据集上有模型具有较高的拟合精度，但是在实际应用中往往精度不佳。因此，如何改善基于神经网络的功率放大器模型的泛化能力是一个亟待解决的问题。
发明内容
本申请实施例提供一种功率放大器模型的获取方法、装置及功率放大器模型、存储介质及程序产品，旨在提升基于神经网络的功率放大器模型的泛化能力。
第一方面，本申请实施例提供一种功率放大器模型的获取方法，所述方法包括：获取功率放大器的初始子模型、标签数据与输入数据；根据所述标签数据与所述输入数据，对所述初始子模型进行迭代训练，直到达到迭代停止条件；并在每次迭代训练完成后，得到一个目标子模型；根据至少一个所述目标子模型得到功率放大器模型。
第二方面，本申请实施例提供功率放大器模型的获取装置，包括：存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序；所述处理器执行所述计算机程序时实现如第一方面所述的功率放大器模型的获取方法。
第三方面，本申请实施例提供功率放大器模型，所述功率放大器模型根据第一方面所述的功率放大器模型的获取方法得到。
第四方面，本申请实施例提供一种计算机可读存储介质，包括：所述计算机可读存储介质存储有计算机可执行指令，所述计算机可执行指令用于执行如第一方面所述的功率放大器模型的获取方法。
第五方面，本申请实施例提供一种计算机程序产品，包括计算机程序或计算机指令，所述计算机程序或所述计算机指令存储在计算机可读存储介质中，计算机设备的处理器从所述计算机可读存储介质读取所述计算机程序或所述计算机指令，所述处理器执行所述计算机程序或所述计算机指令，使得所述计算机设备执行如第一方面所述的功率放大器模型的获取方 法。
根据本申请实施例提供的功率放大器模型的获取方法、装置及功率放大器模型、存储介质及程序产品，通过对初始子模型进行多次迭代训练，得到多个目标子模型，并将得到的多个目标子模型构成的模型作为最终的功率放大器模型，通过优化模型结构与训练过程，提升了功率放大器模型的泛化能力以及模型的预测精度。
附图说明
图1是本申请一实施例中功率放大器的建模方法的应用环境图；
图2是本申请一实施例提供的功率放大器建模的原理图；
图3是本申请一实施例提供的功率放大器模型的获取方法的流程图；
图4是本申请一实施例提供的计算历史时刻输入数据的优先级的流程图；
图5是本申请另一实施例提供的计算历史时刻输入数据的优先级的流程图；
图6是本申请实施例提供的获取预训练的神经网络模型的流程图；
图7是本申请实施例提供的选择目标子模型构建功率放大器模型的流程图；
图8是本申请一示例提供的功率放大器模型结构示意图；
图9是本申请另一示例提供的功率放大器模型结构示意图；
图10是本申请另一示例提供的功率放大器模型结构示意图；
图11是本申请一示例提供的初始子模型的模型结构示意图；
图12(a)是本申请一示例初始子模型中一个单元的模型结构图；
图12(b)是本申请一示例初始子模型中一个单元的模型结构图；
图12(c)是本申请一示例初始子模型中一个单元的模型结构图；
图12(d)是本申请一示例初始子模型中一个单元的模型结构图；
图13是本申请一实施例提供的功率放大器模型结构图；
图14是本申请一实施例提供的功率放大器模型结构图；
图15是本申请一实施例提供的一种功率放大器模型的获取装置结构示意图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本申请，并不用于限定本申请。
需要说明的是，虽然在装置示意图中进行了功能模块划分，在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于装置中的模块划分，或流程图中的顺序执行所示出或描述的步骤。说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。
本申请实施例中，“进一步地”、“示例性地”或者“可选地”等词用于表示作为例子、例证或说明，不应被解释为比其它实施例或设计方案更优选或更具有优势。使用“进一步地”、“示例性地”或者“可选地”等词旨在以具体方式呈现相关概念。
射频功率放大器是无线通信系统中的重要组件，它的作用是把无线信号的功率放大到目标值然后馈入天线。射频功率放大器具非线性和记忆性这两个显著的特性。非线性主要表现在功率放大器对输入信号的增益值随着输入信号功率的增大而发生变化，而非常数。非线性会导致信号的星座图发生畸变，特别是对于正交频分复用(Orthogonal Frequency Division Multiplexing，OFDM)这种高峰均比(Peak to Average Power Ratio，PAPR)的信号。记忆性是指功率放大器的输出不仅和当前时刻的输入有关，还和历史时刻的输入信号有关。记忆性的影响和信号的带宽有关，信号的带宽越大，记忆性造成的影响越大，越不能忽略。
功率放大器的行为建模是功率放大器相关研究工作的基础，例如数字预失真(Digital Pre-Distortion，DPD)。功率放大器建模具体是对功率放大器的非线性和记忆性进行模拟，以建立功率放大器输入和输出之间的数学关系。
相关技术中，用于功率放大器建模的模型大体上可以分为两类，一类是传统模型，例如基于volterra模型的MP模型和GMP模型；另一类是基于神经网络的模型，例如基于MLP网络的RVTDNN模型和基于RNN网络的模型等。传统的功率放大器模型在设计的时候充分考虑了功率放大器的特性，例如模型中的高阶项就对应功率放大器产物中的交调分量，因此它和实际功率放大器的匹配度很好，具有良好的泛化能力。随着功率放大器技术的进步，功率放大器的失真原理变得复杂，传统模型的建模精度大幅下降，研究人员开始把注意力转向基于神经网络的方法。与传统功率放大器模型不同，神经网络是一种通用的模型，功率放大器建模只是它的一种应用场景。由于模型的设计没有考虑功率放大器的失真特性，基于神经网络的模型普遍存在泛化能力差的缺陷，即在训练集上有较高的拟合精度，但是在测试集上拟合精度有较大衰减，特别是当训练集样本数较小的时候。
基于此，本申请实施例提供了一种功率放大器模型的获取方法、装置及功率放大器模型、存储介质及程序产品，通过对初始子模型进行多次迭代训练，得到多个目标子模型，并将得到的多个目标子模型构成的模型作为最终的功率放大器模型，通过优化模型结构与训练过程，提升了功率放大器模型的泛化能力以及模型的预测精度。
本申请提供的功率放大器模型的获取方法，可以应用在图1所述的应用环境中。其中，电子设备100获取功率放大器200的输入输出采样值，然后对功率放大器的输入输出特征进行建模，获得功率放大器模型300。其中上述电子设备可以基站中的预失真模块，也可以是雷达系统中的信号处理单元，本申请实施例在此不做限定。上述功率放大器可以是基站中的功率放大器组件，也可以是雷达系统中的功率放大器单元，也可以是终端设备中的功率放大器器件等，本申请实施例在此不做限定。
图2为本申请一实施例提供的功率放大器建模的原理图。如图所示，假设功率放大器的输入和输出信号分别为x和y，对其进行建模的原理如图2所示，其目标是使得建立的功率放大器模型的输出和PA的输出y之间的误差e最小。
图3为本申请一实施例提供的功率放大器模型的获取方法的流程图。如图3所示，功率放大器模型的获取方法可以包括但不限于步骤S1000、步骤S2000以及步骤S3000。
步骤S1000：获取功率放大器的初始子模型、标签数据与输入数据。
在一实施例中，功率放大器的初始子模型为一种神经网络模型，标签数据为功率放大器 实际输出信号与功率放大器初始子模型输出信号之间的误差，输入数据为当前时刻的输入信号与历史时刻的输入信号。由于考虑到历史时刻的输入信号对当前时刻的输出产生影响，因此，本申请的实施例在模型输入数据的获取时，会获取历史时刻的输入信号并作为输入数据进行模型训练。在功率放大器技术领域，常使用记忆深度来表示当前的输出信号与多久前的输入信号有关。
在另一实施例中，功率放大器的初始子模型为多个神经网络模型构成的模型，多个神经网络模型构成的初始子模型对数据的拟合可能更具有优势。
可以理解的是，初始子模型可以为一种神经网络模型，也可以是多个神经网络模型组成的模型，可以根据应用场景的不同需要而选取不同的初始子模型。
步骤S2000：根据标签数据与输入数据，对初始子模型进行迭代训练，直到达到迭代停止条件；并在每次迭代训练完成后，得到一个目标子模型。
在一实施例中，使用标签数据与输入数据对初始子模型进行迭代训练，当达到迭代停止条件时，就停止本次迭代训练，并获得一个训练后的目标子模型。在每次迭代训练后，根据目标子模型的输出，更新标签数据，得到新的标签数据，新的标签数据用于下一次子模型迭代训练。
在一实施例中，迭代停止条件为预设的迭代次数，因此，在当前迭代次数小于预设的迭代次数的情况下，将当前时刻输入数据与至少一个历史时刻输入数据输入至初始子模型进行持续的迭代训练，并在每次迭代训练后，生成一个目标子模型。多次迭代训练得到的多个目标子模型有机会构成最终的功率放大器模型。
在一实施例中，迭代停止条件为预设的迭代次数，因此，在当前迭代次数小于预设的迭代次数的情况下，将当前时刻输入数据与所有历史时刻输入数据输入至初始子模型进行持续的迭代训练，并在每次迭代训练后，生成一个目标子模型。多次迭代训练得到的多个目标子模型有机会构成最终的功率放大器模型。
在一实施例中，迭代停止条件为预设的迭代次数，因此，在当前迭代次数小于预设的迭代次数的情况下，每次迭代训练过程中，将当前时刻输入数据与部分所述历史时刻输入数据输入至初始子模型，直至达到预设的迭代次数，就停止迭代训练，并在每次迭代训练后，生成一个目标子模型。多次迭代训练得到的多个目标子模型有机会构成最终的功率放大器模型。
在一实施例中，迭代停止条件为所有历史时刻输入数据均参与模型迭代训练，因此，在当前迭代训练过程中，根据预设的优先级计算条件，生成每个历史时刻输入数据的优先级，根据当前子模型迭代训练次数与历史时刻输入数据的优先级，构建历史时刻输入数据集合，在当前迭代训练的输入数据未达到迭代停止条件的情况下，将当前时刻输入数据与历史时刻输入数据集合输入至初始子模型进行迭代训练。需要说明的是，为了构建历史时刻输入数据集合，需要根据历史时刻输入数据的优先级对历史时刻输入数据进行排序，得到排序后的历史时刻输入数据，再在所述排序后的历史时刻输入数据中，依次选取目标数量的历史时刻输入数据，构建历史时刻输入数据集合，目标数量等于当前子模型迭代训练的次数，可以理解的是，随着子模型迭代训练的次数的增加，越来越多的历史时刻输入数据被选入构成历史时刻输入数据集合，直至所有的历史时刻数据均被选入构成历史时刻输入数据集合，此时结束 子模型的迭代训练。而由于子模型的首次迭代训练(第0次迭代)仅需要输入当前时刻数据作为训练数据，因此，在本实施例中，假如记忆深度为Q，子模型的迭代训练过程要进行Q+1次，最终生成Q+1个目标子模型。
可以理解的是，在一些实施例中，可以采用其他方式构建历史时刻输入数据集合，例如在每次迭代训练时，选取两个或者更多个排名靠前的历史时刻输入数据加入到前一次迭代训练构建的历史时刻输入数据集合中，形成新的历史时刻输入数据集合，通过这种方法，能够减少子模型迭代训练的总次数，提高模型训练的效率。
可以理解的是，对历史时刻输入数据进行优先级的排序是为了找到对当前模型输出影响最大历史时刻输入数据，因此有多种方式可以对历史时刻输入数据的优先级进行计算。下面，将提供几个实施例进行详细说明。
图4是本申请一实施例提供的计算历史时刻输入数据的优先级的流程图。在本实施例中，计算历史时刻输入数据至少包括步骤S2110、S2120以及S2130。
步骤S2110：获取当前子模型迭代训练的误差值。
在一实施例中，计算当前迭代训练过程中的残余信号，残余信号就是功率放大器实际输出与模型预测输出之间的差值。
可以理解的是，其他能够表示功率放大器实际输出与模型预测输出之间的差距的方式也可以被用来计算相关性。本申请实施例在此不做限制。
步骤S2120：计算误差值与历史时刻输入数据之间的相关性。
在一实施例中，根据相关度计算公式计算误差值与历史时刻输入数据之间的相关度，并得到相关度值。
可以理解的是，能够计算误差值与历史时刻输入数据之间的相关度的方法均可以用于本申请实施例步骤S2120的相关度计算中。本申请实施例在此不做限制。
步骤S2130：根据相关性，得到历史时刻输入数据的优先级。
在一实施例中，对得到的相关度值进行排序，相关度值越大，表示相关度越高，说明对应的历史时刻输入数据对模型当前输出的影响越大，该历史时刻输入数据的优先级就会更高。将优先级高的历史时刻输入数据优先构建历史时刻输入数据集合，能够提高最终功率放大器模型的生成效率与预测精度。
图4的实施例按照历史时刻输入数据的优先级，依次构建具有不同历史时刻输入数据的多个历史时刻输入数据集合，使用历史时刻输入数据集合进行子模型的迭代训练，生成多个目标子模型，能够有效提高模型的训练效率与模型在实际应用场景下的泛化能力。
图5是本申请另一实施例提供的计算历史时刻输入数据的优先级的流程图。在本实施例中，计算历史时刻输入数据至少包括步骤S2210、S2220以及S2230。
步骤S2210：获取预训练的神经网络模型。
在一实施例中，获取经过预训练的神经网络模型，作为对历史时刻输入数据拟合精度的预测模型。预训练的神经网络模型是通过训练数据训练后的模型，能够在实际预测时，生成对应于输入的历史时刻输入数据的拟合精度。
图6为本申请实施例提供的获取预训练的神经网络模型的流程图。获取预训练的神经网 络模型具体包括步骤S2211和S2212。
步骤S2211：根据当前时刻输入数据与至少一个历史时刻输入数据，通过组合方式，构建多个训练数据集合。
在一实施例中，由于模型是为了对历史时刻输入数据的拟合精度进行预测，因此，在训练此模型时，使用的训练数据中的输入数据为当前时刻输入数据与历史时刻输入数据的组合，训练数据的标签数据为实际输出与预测输出的误差值。
在一实施例中，对历史时刻输入数据的优先级进行排序，并选取排名靠前的历史时刻输入数据与当前时刻输入数据一起构建输入数据，将构建的输入数据称为临时队列。在本实施例中，构建第一个临时队列时，临时队列中包含当前时刻的输入数据与优先级排名最高的历史时刻输入数据；构建第二个临时队列时，临时队列中包含当前时刻的输入数据与优先级排名前两位的历史时刻输入数据，以此类推，直至所有的历史时刻输入数据均被选入构建临时队列。
可以理解的是，在此步骤中是对模型的预训练，因此可以尽可能多的构建临时队列，以获得更多的训练数据，这样在模型的训练过程中，能够生成预测结果更加精准的模型。但是，在一些应用场景下，如果局限于计算能力或者为了防止模型的过拟合，可以选择部分的历史时刻输入数据与当前输入数据构成训练数据，这些被选择的历史时刻输入数据可以是随机选择的。
步骤S2212：分别将多个训练数据集合输入至神经网络模型，通过标签数据对神经网络模型进行训练，得到预训练的神经网络模型。
在一实施例中，在选取好神经网络模型后，可以对模型进行训练，训练的步骤与常用神经网络的训练相同，本申请在此不做赘述。
可以理解的是，对于此步骤中神经网络模型的选择可以根据实际需求调整，而对于神经网络模型的训练也可以采用常用的有监督学习方式。本申请实施例在此不做限制。
步骤S2220：将历史时刻输入数据输入至预训练的神经网络模型进行拟合，得到对应于历史时刻输入数据的拟合精度值。
在一实施例中，将历史时刻输入数据输入至经过训练的神经网络模型，能够预测历史时刻输入数据的拟合精度值。
步骤S2230：根据拟合精度值，得到历史时刻输入数据的优先级。
在一实施例中，拟合精度值越高，则该历史时刻输入数据对应的优先级就越高，拟合精度值越低，则该历史时刻输入数据对应的优先级就越低。
可以理解的是，在实际应用中，当预测的拟合精度值超过预设的阈值后，可以不对这些拟合精度超过阈值的历史时刻输入数据再进行排序了，这些具有较高拟合精度的历史时刻输入数据在历史时刻输入数据集合时具有同等的地位。
步骤S3000：根据至少一个目标子模型得到功率放大器模型。
在一实施例中，子模型迭代训练生成的所有目标子模型均用于构建最终的功率放大器模型。通过这种方式获得的功率放大器模型由于包含了所有历史时刻输入数据对模型输出的影响，因此在一些应用场景下，可能预测精度较高。
在另一实施例中，仅选取迭代训练所获得的部分目标子模型构建最终的功率放大器模型。其原因是，在在一些应用场景下，功率放大器虽然存在记忆性，但是实际的输出仅和几个特定时刻点的历史输入相关。因此，可以通过在子模型迭代训练过程中对某一个或多个统计维度设定阈值，例如均方差(Mean squared error，MSE)或标准均方差(normalized mean squared error，NMSE)，而网络子模型迭代训练至收敛后该统计量劣于此阈值，则可以认为这个目标子模型对于构建最终的功率放大器模型是没有帮助的，因此丢弃掉这些不满足阈值要求的一个或多个目标子模型。
图7是本申请实施例提供的选择目标子模型构建功率放大器模型的流程图。如图所示，至少包括步骤S3100、S3200、S3300以及S3400。
步骤S3100：获取预设统计维度和与预设统计维度对应的预设统计维度阈值。
在一实施例中，选取均方差作为统计维度，并提前预设均方差阈值。如果目标子模型的的均方差超过预设的均方差阈值，则说明该子模型训练效果达不到预期，则后续可能放弃掉该目标子模型，而选取其他均方差没有超过预设的均方差阈值的目标子模型构建最终的功率放大器模型。
步骤S3200：获取每个目标子模型对应于预设统计维度的统计值。
在一实施例中，获取每个目标子模型的均方差。
步骤S3300：将统计值优于预设统计维度阈值的目标子模型确定为功率放大器子模型。
在一实施例中，将均方差小于或等于均方差阈值的目标子模型选取出来，这些满足要求的目标子模型被最终用于构建功率放大器模型。
在另一实施例中，虽然所有的目标子模型的均方差均小于或等于均方差阈值，但是为了实现简化模型的目的，可以选择预设数量的均方差相对更小的多个目标子模型进行功率放大器模型的构建。
在另一实施例中，也采取使用部分目标子模型构建最终的功率放大器模型的方式，但是与上一实施例不同之处在于，本实施例中不通过对统计维度预先设置阈值的方式进行模型的筛选，而是直接预先设定最终功率放大器模型的复杂度，并且依据设定的复杂度，对目标子模型的收敛程度进行排序，丢弃掉一个或多个目标子模型，使剩余的目标子模型构成的功率放大器模型满足预设的模型复杂度。
可以理解的是，本申请实施例对于目标子模型的选择方式可以有多种，不局限于上述实施例提供的几种方式。
步骤S3400：将统计值优于预设统计维度阈值的目标子模型确定为功率放大器子模型。
在一实施例中，根据均方差统计值，对功率放大器子模型进行排序，得到排序顺序，将功率放大器子模型按照排序顺序构成的模型确定为功率放大器模型。
因此，仅采用部分目标子模型构建最终的功率放大器模型的实施例能够通过减少模型所使用的子模型的个数，从而减少模型的参数，降低模型的复杂度，从而提升模型的泛化能力。
下面将通过五个示例详细说明本申请实施例提供的功率放大器模型的获取方法的应用过程。
示例一：
设定功率放大器模型的记忆深度为Q，则功率放大器模型的输入为x(n)，x(n-1)，……，x(n-Q)，其中x(n)表示当前时刻输入数据，其余为记忆项，即历史时刻输入数据；初始化迭代计数器iter＝1；初始化模型输出
步骤S101：获取初始子模型，即，初始化子模型NNiter。
步骤S102：生成初始子模型NNiter的输入数据Xiter和标签数据Yiter，其中，输入数据Xiter和标签数据Yiter的生成方式如下：
下面更为具体地说明训练过程。
例如，在第一次迭代过程中，使用当前时刻输入数据x(n)进行训练；第二次迭代过程中，使用当前时刻输入数据x(n)与记忆深度为1的历史时刻输入数据x(n-1)进行训练；第三次迭代过程中，使用当前时刻输入数据x(n)、记忆深度为1的历史时刻输入数据x(n-1)以及记忆深度为2的历史时刻输入数据x(n-2)进行训练，以此类推。
步骤S103：训练初始子模型至收敛。
步骤S104：根据生成的目标子模型，更新功率放大器模型的输出
步骤S105：在本示例中，迭代训练停止条件为预设的迭代次数阈值M。判断是否满足迭代训练停止条件，如果满足则停止迭代，如果不满足则更新迭代计数器iter＝iter+1，且返回步骤S101。
在迭代训练M次后停止迭代，算法流程结束后得到，NN1，NN2，…，NNM，这M个目标子模型，这M个目标子模型共同构成了本申请示例提出的功率放大器模型，该模型的原理图如图8所示。
示例二：
设定功率放大器模型的记忆深度为Q；初始化队列S1＝{0}，队列S2＝{1,2,...,Q}；初始化迭代计数器iter＝1；初始化模型输出
示例二与示例一的相同之处在于采用相同的初始子模型，不同之处在于示例二中的子模型训练方式不同，具体如下：
步骤S201：获取初始子模型，即，初始化子模型NNiter。
步骤S202：生成子模型神经网络NNiter的输入数据X和标签数据Y，X和Y的生成方式如下，；
其中表示队列S1中第i个元素。
步骤S203：训练子模型神经网络NNiter至收敛。
步骤S204：根据生成的目标子模型，更新功率放大器模型的输出
步骤S205：判断队列S2是否为空，如果为空则流程结束。
需要说明的是，在首次迭代训练中，仅使用当前时刻输入数据进行模型训练，因此，队列S2一定不为空。
步骤S206：计算队列S2内的元素的优先级顺序。
在示例中，采用相关性度量方法对队列S2中的历史时刻输入数据的优先级进行计算。具体计算过程如下：
子步骤1：计算当前迭代的残余信号r，
子步骤2：计算队列S2中第i个元素的相关度量
子步骤3：根据相关性度量进行优先级排序，相关度量越大优先级越高。
可以理解的是，采用相关性度量方法进行优先级计算的计算复杂度较低，能够较为快速地生成优先级值，进而得到优先级顺序。
步骤S207：将队列S2中优先级最高的元素加入到队列S1的末尾，并从队列S2中将该元素删除。
需要说明的是，每次将一个队列S2中的历史时刻输入数据加入到队列S1中，会导致队列S2中元素个数逐渐减少，直至队列S2为空，而队列S1中元素个数会逐渐增加，也就是用于子模型迭代训练的输入数据个数会逐渐增加，逐渐由仅采用当前时刻输入数据进行训练变化为将越来越多的历史时刻输入数据也一起输入参与训练过程。
可以理解的是，本示例中，每次将一个历史时刻输入数据加入到当前时刻输入数据中，从而构成本次迭代训练的输入数据，而在其他示例中，如果记忆深度过大，可以每次将两个或两个以上的历史时刻输入数据加入到当前时刻输入数据中，从而节约模型训练的时间，提高模型训练的效率。
步骤S208：更新迭代计数器iter＝iter+1，且返回步骤S201。
在Q+1次后停止迭代训练，算法流程结束后得到NN1，NN2，…，NNQ+1这Q+1个目标子模型，这Q+1个目标子模型共同构成了本申请示例提出的功率放大器模型，该模型的原理图如图9所示。
图中的优先级排序模块就是按照队列S1中元素的顺序对x(n)，x(n-1)，…，x(n-Q)进行排序输出，即
示例三：
设定功率放大器模型的记忆深度为Q；初始化队列S1＝{0}，队列S2＝{1,2,...,Q}；初始化迭代计数器iter＝1；初始化模型输出
示例三与示例二的相同之处在于采用相同的初始子模型与训练思路，不同之处在于示例三中历史时刻输入数据的优先级计算方法不同，具体如下：
步骤S301：获取初始子模型，即，初始化子模型NNiter。
步骤S302：生成子模型神经网络NNiter的输入数据X和标签数据Y，X和Y的生成方式如下，；
其中表示队列S1中第i个元素。
步骤S303：训练子模型神经网络NNiter至收敛。
步骤S304：根据生成的目标子模型，更新功率放大器模型的输出
步骤S305：判断队列S2是否为空，如果为空则流程结束。
需要说明的是，在首次迭代训练中，仅使用当前时刻输入数据进行模型训练，因此，队列S2一定不为空。
步骤S306：计算队列S2内的元素的优先级顺序。
在示例中，采用相关性度量方法对队列S2中的历史时刻输入数据的优先级进行计算。具体计算过程如下：
子步骤1：生成临时队列Ω， 表示队列S2中第i个元素；
子步骤2：初始化神经网络NNtmp。
需要说明的是，神经网络NNtmp可以是符合需求的任意种类的神经网络。
子步骤3：生成神经网络NNtmp的输入数据X和标签数据Y，X和Y的生成方式如下：
其中Ωi表示队列Ω中第i个元素。
需要说明的是，生成临时队列的目的是对神经网络NNtmp进行训练，使经过训练后的模型NNtmp能够对应用过程中的历史时刻输入数据的拟合精度进行准确预测，从而得出哪些历史时刻输入数据对输出结果的影响更大，进而赋予其更高的优先级。
子步骤4：利用上述输入数据与标签数据，训练神经网络NNtmp。
子步骤5：计算NNtmp的拟合精度
需要说明的是，在模型完成训练后，就得到了一个预训练的神经网络模型。将历史时刻输入数据输入至这个预训练的神经网络模型，对其拟合精度进行预测。
子步骤6：根据拟合精度进行优先级排序，拟合精度越高优先级越高。
步骤S307：将队列S2中优先级最高的元素加入到队列S1的末尾，并从队列S2中将该元素删除。
需要说明的是，每次将一个队列S2中的历史时刻输入数据加入到队列S1中，会导致队列S2中元素个数逐渐减少，直至队列S2为空，而队列S1中元素个数会逐渐增加，也就是用于子模型迭代训练的输入数据个数会逐渐增加，逐渐由仅采用当前时刻输入数据进行训练变化为将越来越多的历史时刻输入数据也一起输入参与训练过程。
可以理解的是，本示例中，每次将一个历史时刻输入数据加入到当前时刻输入数据中，从而构成本次迭代训练的输入数据，而在其他示例中，如果记忆深度过大，可以每次将两个或两个以上的历史时刻输入数据加入到当前时刻输入数据中，从而节约模型训练的时间，提 高模型训练的效率。
步骤S308：更新迭代计数器iter＝iter+1，且返回步骤S301。
在Q+1次后停止迭代训练，算法流程结束后得到NN1，NN2，…，NNQ+1这Q+1个目标子模型，这Q+1个目标子模型共同构成了本申请示例提出的功率放大器模型，该模型的原理图如图9所示。
图中的优先级排序模块就是按照队列S1中元素的顺序对x(n)，x(n-1)，…，x(n-Q)进行排序输出，即
示例四：
在一些情况下，需要对模型的复杂度进行简化，从而使其适应更多的应用场景。因此，在前述示例一、二、三中得到的多个目标子模型可能不能全部被选择用来构建最终的功率放大器模型，此时，就需要对这些目标子模型进行选择。
具体方法为，对神经网络NNiter训练中的某一个统计维度设定阈值，这个统计维度可以是均方差MSE或标准均方差NMSE。如果网络NNiter训练至收敛后该统计量劣于此阈值，则可以认为在NNiter对于构建Y是没有帮助的，因此NNiter是不必要的，NNiter不参与最终功率放大器模型的构建。
示例四仍然可以采用示例一中所选取的初始子模型，但是不同之处在于，在训练过程中对目标子模型进行了筛选，具体如下：
设定功率放大器模型的记忆深度为Q；设定统计量阈值为初始化队列S1＝{0}，队列S2＝{1,2,...,Q}；初始化迭代计数器iter＝1；初始化模型输出
步骤S401：初始化神经网络NNiter。
步骤S402：生成神经网络NNiter的输入数据X和标签数据Y，X和Y的生成方式如下：
步骤S403：训练神经网络NNiter至收敛。
步骤S404：判断NNiter的统计量和阈值的关系，如果优于设定的阈值则认定NNiter有效，更新功率放大器模型的输出如果劣于设定的阈值则认定NNiter无效，NNiter不参与最终功率放大器模型的构建。
步骤S405：判断队列S2是否为空，如果为空则流程结束。
步骤S406：计算队列S2内的元素的优先级顺序。
步骤S407：将队列S2中优先级最高的元素加入到队列S1的末尾，并从队列S2中将该元素删除。
步骤S408：更新迭代计数器iter＝iter+1，且返回步骤S401。
假设在一次功率放大器建模中，记忆深度Q设定为4，目标统计量为标准均方差NMSE， 标准均方差NMSE的计算式为标准均方差在训练完成后队列S1的元素为{0,1,4,2,3}，NN2和NN4在训练完成后统计指标标准均方差NMSE高于设定的阈值，则认为NN2和NN4无效，不参与最终功率放大器模型的构建，最终的功率放大器模型如图10所示。
本示例通过这种方式可以减少模型所使用的目标子模型的数量，从而减少模型的参数，降低模型的复杂度。
示例五：
在本示例中，提供一种初始子模型NNiter的模型结构。图11为本申请示例提供的初始子模型的模型结构示意图。
如图所示，单元1的主体结构是一个复数多层感知器(Multilayer Perceptron，MLP)网络，单元1的输出和输入的绝对值构成了该MLP网络的输入。单元2是将单元1的输出和X进行相乘再求和。单元3是对单元1的输出进行加权求和，其中权值系数w1,w2,...,witer是该网络的参数，需要训练得到。单元4是一个线性模型，实质是对X加权求和，其中权值系数θ1,θ2,...,θiter是该网络的参数，需要训练得到。其中
为了进一步说明初始子模型中每个单元的具体模型结构，提供图12(a)-(d)是本申请示例初始子模型中每个单元的模型结构图。
需要说明的是，本示例提供的初始子模型结构仅作为样例，本领域技术人员可以根据实际需求选择合适的初始子模型结构进行训练。
本申请示例提供的功率放大器模型的获取方法通过优化模型结构与训练过程，提升了功率放大器模型的泛化能力以及模型的预测精度。
在一实施例中，根据本申请实施例提供的功率放大器模型的获取方法，得到了最终的功率放大器模型。
下面对通过本申请实施例提供的功率放大器模型的获取方法得到的最终的功率放大器模型进行性能的验证，验证过程如下：
步骤S501：从功率放大器采集一组输入输出数据x和y。
步骤S502：分别使用传统的基于神经网络的功率放大器建模方法以及本申请实施例得到的功率放大器建模方法从这组数据中提取功率放大器模型。
步骤S503：对步骤S502得到的功率放大器模型计算得到数字预失真DPD模型。
步骤S504：将x通过步骤S503得到的DPD模型得到信号z，再将信号z通过功率放大器得到信号
步骤S505：通过x和计算DPD性能统计量其中G是功率放大器的增益。
通过比较两种功率放大器建模方法获得的统计量P来衡量两种建模方法的优劣，P越小则功率放大器模型越接近真实的功率放大器。
传统的基于神经网络的功率放大器建模和本申请实施例提出的功率放大器建模方法都使用示例五中的神经网络模型，区别在于传统模型只包含一个神经网络，输入数量为Q+1，本申请的提出的功率放大器模型包含Q+1个神经网络子模型，输入数量分别是1，2，……，Q+1。本实施例中功率放大器模型的记忆深度Q设置为9。
首先使用传统的基于神经网络的功率放大器建模的方法，分别对20M和40M两种带宽的信号进行测试，按照本实施例实际的测试方法，最终的统计量如下表1所示。
表1
然后使用本申请实施例提出的功率放大器建模的方法，分别对20M和40M两种带宽的信号进行测试，按照本实施例实际的测试方法，根据示例四中，设定统计量阈值为-1dB。最终的统计量如下表2所示。
表2
20M带宽下，模型训练至第4个迭代之后的神经网络都不能满足阈值的要求，对构建功率放大器模型没有帮助。训练完成后队列S1＝{0,5,9,1,3,6,2,8,4,7}，则最终的功率放大器模型如图13所示。
40M带宽下，模型训练至第5个迭代之后的神经网络都不能满足阈值的要求，对构建功率放大器模型没有帮助。训练完成后队列S1＝{0,2,6,9,8,4,5,7,3}，则最终的功率放大器模型如图14所示。
对比两种功率放大器建模的方法可以看出，两种方法的功率放大器拟合精度差别很小，但是在将模型实际用在数字预失真DPD测试的时候差距巨大，本申请实施例提出的功率放大器建模方法性能明显更优，这说明了本申请实施例提出建模方法所建立的PA模型更接近真实的功率放大器，具有更好的泛化能力。
图15是本申请一实施例提供的一种功率放大器模型的获取装置结构示意图。如图15所示，该设备包括存储器1100、处理器1200。存储器1100、处理器1200的数量可以是一个或多个，图15中以一个存储器1100和一个处理器1200为例；设备中的存储器1100和处理器1200可以通过总线或其他方式连接，图15中以通过总线连接为例。
存储器1100作为一种计算机可读存储介质，可用于存储软件程序、计算机可执行程序以及模块，如本申请任一实施例提供的资源确定方法对应的程序指令/模块。处理器1200通过运行存储在存储器1100中的软件程序、指令以及模块实现上述功率放大器模型的获取方法。
存储器1100可主要包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、至少一个功能所需的应用程序。此外，存储器1100可以包括高速随机存取存储器，还可以包括非易失性存储器，例如至少一个磁盘存储器件、闪存器件或其他非易失性固态存储器件。在一些实例中，存储器1100可进一步包括相对于处理器1200远程设置的存储器，这些远程存储器可以通过网络连接至设备。上述网络的实例包括但不限于互联网、企业内部网、局域网、移动通信网及其组合。
本申请一实施例还提供了一种计算机可读存储介质，存储有计算机可执行指令，该计算机可执行指令用于执行如本申请任一实施例提供的功率放大器模型的获取方法。
本申请一实施例还提供了一种计算机程序产品，包括计算机程序或计算机指令，计算机程序或计算机指令存储在计算机可读存储介质中，计算机设备的处理器从计算机可读存储介质读取计算机程序或所述计算机指令，处理器执行所述计算机程序或所述计算机指令，使得计算机设备执行如本申请任一实施例提供的功率放大器模型的获取方法。
本申请实施例描述的系统架构以及应用场景是为了更加清楚的说明本申请实施例的技术方案，并不构成对于本申请实施例提供的技术方案的限定，本领域技术人员可知，随着系统架构的演变和新应用场景的出现，本申请实施例提供的技术方案对于类似的技术问题，同样适用。
本领域普通技术人员可以理解，上文中所公开方法中的全部或某些步骤、系统、设备中的功能模块/单元可以被实施为软件、固件、硬件及其适当的组合。
在硬件实施方式中，在以上描述中提及的功能模块/单元之间的划分不一定对应于物理组件的划分；例如，一个物理组件可以具有多个功能，或者一个功能或步骤可以由若干物理组件合作执行。某些物理组件或所有物理组件可以被实施为由处理器，如中央处理器、数字信号处理器或微处理器执行的软件，或者被实施为硬件，或者被实施为集成电路，如专用集成电路。这样的软件可以分布在计算机可读介质上，计算机可读介质可以包括计算机存储介质(或非暂时性介质)和通信介质(或暂时性介质)。如本领域普通技术人员公知的，术语计算机存储介质包括在用于存储信息(诸如计算机可读指令、数据结构、程序模块或其他数据)的任何方法或技术中实施的易失性和非易失性、可移除和不可移除介质。计算机存储介质包括但不限于RAM、ROM、EEPROM、闪存或其他存储器技术、CD-ROM、数字多功能盘(DVD)或其他光盘存储、磁盒、磁带、磁盘存储或其他磁存储装置、或者可以用于存储期望的信息并且可以被计算机访问的任何其他的介质。此外，本领域普通技术人员公知的是，通信介质通常包含计算机可读指令、数据结构、程序模块或者诸如载波或其他传输机制之类的调制数据信号中的其他数据，并且可包括任何信息递送介质。
在本说明书中使用的术语“部件”、“模块”、“系统”等用于表示计算机相关的实体、硬件、固件、硬件和软件的组合、软件、或执行中的软件。例如，部件可以是但不限于，在处理器上运行的进程、处理器、对象、可执行文件、执行线程、程序或计算机。通过图示，在计算 设备上运行的应用和计算设备都可以是部件。一个或多个部件可驻留在进程或执行线程中，部件可位于一个计算机上或分布在2个或更多个计算机之间。此外，这些部件可从在上面存储有各种数据结构的各种计算机可读介质执行。部件可例如根据具有一个或多个数据分组(例如来自于自与本地系统、分布式系统或网络间的另一部件交互的二个部件的数据，例如通过信号与其它系统交互的互联网)的信号通过本地或远程进程来通信。

Claims (15)

  1. 一种功率放大器模型的获取方法，包括：
    获取功率放大器的初始子模型、标签数据与输入数据；
    根据所述标签数据与所述输入数据，对所述初始子模型进行迭代训练，直到达到迭代停止条件；并在每次迭代训练完成后，得到一个目标子模型；
    根据至少一个所述目标子模型得到功率放大器模型。
  2. 根据权利要求1所述的功率放大器模型的获取方法，其中，所述在每次迭代训练完成后，所述方法还包括：
    根据所述目标子模型，更新所述标签数据；其中，更新后的所述标签数据用于下一次子模型迭代训练。
  3. 根据权利要求2所述的功率放大器模型的获取方法，其中，所述输入数据包括当前时刻输入数据与多个历史时刻输入数据，所述迭代停止条件为预设的迭代次数；
    所述对所述初始子模型进行迭代训练，包括：
    在当前迭代次数小于所述预设的迭代次数的情况下，将所述当前时刻输入数据与至少一个所述历史时刻输入数据输入至所述初始子模型进行迭代训练。
  4. 根据权利要求2所述的功率放大器模型的获取方法，其中，所述迭代停止条件为所有历史时刻输入数据均参与模型迭代训练；
    所述对所述初始子模型进行迭代训练，包括：
    根据预设的优先级计算条件，生成每个所述历史时刻输入数据的优先级；
    根据当前子模型迭代训练次数与所述历史时刻输入数据的优先级，构建历史时刻输入数据集合；
    在当前迭代训练的输入数据未达到所述迭代停止条件的情况下，将所述当前时刻输入数据与所述历史时刻输入数据集合输入至所述初始子模型进行迭代训练。
  5. 根据权利要求4所述的功率放大器模型的获取方法，其中，所述根据当前子模型迭代训练次数与所述历史时刻输入数据的优先级，构建历史时刻输入数据集合，包括：
    根据所述历史时刻输入数据的优先级对所述历史时刻输入数据进行排序，得到排序后的历史时刻输入数据；
    在所述排序后的历史时刻输入数据中，依次选取目标数量的历史时刻输入数据，构建所述历史时刻输入数据集合；其中，所述目标数量等于所述当前子模型迭代训练的次数。
  6. 根据权利要求4或5所述的功率放大器模型的获取方法，其中，所述根据预设的优先级计算条件，生成每个所述历史时刻输入数据的优先级，包括：
    获取当前子模型迭代训练的误差值；
    计算所述误差值与所述历史时刻输入数据之间的相关性；
    根据所述相关性，得到所述历史时刻输入数据的优先级。
  7. 根据权利要求4或5所述的功率放大器模型的获取方法，其中，所述根据预设的优先级计算条件，生成每个所述历史时刻输入数据的优先级，包括：
    获取预训练的神经网络模型；
    将所述历史时刻输入数据输入至所述预训练的神经网络模型进行拟合，得到对应于所述历史时刻输入数据的拟合精度值；
    根据所述拟合精度值，得到所述历史时刻输入数据的优先级。
  8. 根据权利要求7所述的功率放大器模型的获取方法，其中，所述获取预训练的神经网络模型，包括：
    根据所述当前时刻输入数据与至少一个所述历史时刻输入数据，通过组合方式，构建多个训练数据集合；
    分别将多个所述训练数据集合输入至神经网络模型，通过所述标签数据对所述神经网络模型进行训练，得到所述预训练的神经网络模型。
  9. 根据权利要求1所述的功率放大器模型的获取方法，其中，所述根据至少一个所述目标子模型得到功率放大器模型，包括：
    获取预设统计维度和与所述预设统计维度对应的预设统计维度阈值；
    获取每个所述目标子模型对应于所述预设统计维度的统计值；
    将所述统计值优于所述预设统计维度阈值的所述目标子模型确定为功率放大器子模型；
    将至少一个所述功率放大器子模型构成的模型确定为所述功率放大器模型。
  10. 根据权利要求9所述的功率放大器模型的获取方法，其中，所述预设统计维度为标准均方差，所述预设统计维度阈值为标准均方差阈值；
    所述将所述统计值优于所述预设统计维度阈值的所述目标子模型确定为功率放大器子模型，包括：
    将每个所述目标子模型的标准均方差与所述标准均方差阈值进行比较；
    将标准均方差小于或等于所述标准均方差阈值的所述目标子模型作为所述功率放大器子模型。
  11. 根据权利要求9或10所述的功率放大器模型的获取方法，其中，所述将至少一个所述目标子模型构成的模型确定为功率放大器模型，包括：
    根据所述统计值，对所述功率放大器子模型进行排序，得到排序顺序；
    将所述功率放大器子模型按照所述排序顺序构成的模型确定为所述功率放大器模型。
  12. 一种功率放大器模型的获取装置，包括：
    存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序；
    所述处理器执行所述计算机程序时实现如权利要求1至11中任一项所述的功率放大器模型的获取方法。
  13. 一种功率放大器模型，其中，所述功率放大器模型根据权利要求1至11中任一项所述的功率放大器模型的获取方法得到。
  14. 一种计算机可读存储介质，存储有计算机可执行指令，其中，所述计算机可执行指令用于执行如权利要求1至11任一项所述的功率放大器模型的获取方法。
  15. 一种计算机程序产品，包括计算机程序或计算机指令，其中，所述计算机程序或所述计算机指令存储在计算机可读存储介质中，计算机设备的处理器从所述计算机可读存储介质读取所述计算机程序或所述计算机指令，所述处理器执行所述计算机程序或所述计算机指令，使得所述计算机设备执行如权利要求1至11任意一项所述的功率放大器模型的获取方法。
PCT/CN2023/080821 2022-07-26 2023-03-10 功率放大器模型的获取方法、装置及功率放大器模型 Ceased WO2024021621A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
US18/875,106 US20250315727A1 (en) 2022-07-26 2023-03-10 Method and apparatus for acquiring power amplifier model, and power amplifier model
KR1020257000396A KR20250018557A (ko) 2022-07-26 2023-03-10 전력 증폭기 모델의 획득 방법, 장치 및 전력 증폭기 모델
EP23844839.3A EP4546204A4 (en) 2022-07-26 2023-03-10 METHOD AND APPARATUS FOR ACQUIRING POWER AMPLIFIER MODEL, AND POWER AMPLIFIER MODEL

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210885045.6 2022-07-26
CN202210885045.6A CN117540677A (zh) 2022-07-26 2022-07-26 功率放大器模型的获取方法、装置及功率放大器模型

Publications (1)

Publication Number Publication Date
WO2024021621A1 true WO2024021621A1 (zh) 2024-02-01

Family

ID=89705167

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/080821 Ceased WO2024021621A1 (zh) 2022-07-26 2023-03-10 功率放大器模型的获取方法、装置及功率放大器模型

Country Status (5)

Country Link
US (1) US20250315727A1 (zh)
EP (1) EP4546204A4 (zh)
KR (1) KR20250018557A (zh)
CN (1) CN117540677A (zh)
WO (1) WO2024021621A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117970004A (zh) * 2024-02-22 2024-05-03 南京莱芯科技有限公司 功率放大器的测试方法、装置、介质和电子设备
CN118764003A (zh) * 2024-09-09 2024-10-11 中国移动紫金(江苏)创新研究院有限公司 数据预失真处理方法、装置、设备、介质及程序产品
CN118801487A (zh) * 2024-09-14 2024-10-18 内蒙古电力（集团）有限责任公司内蒙古电力经济技术研究院分公司 一种面向新型电力系统的多类型储能资源配置方法
WO2025256215A1 (zh) * 2024-06-11 2025-12-18 华为技术有限公司 一种通信方法及相关装置

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119692271B (zh) * 2024-12-03 2026-02-03 中国科学院自动化研究所 功率放大器工作特性模拟方法、装置、计算机设备及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105224985A (zh) * 2015-09-28 2016-01-06 南京航空航天大学 一种基于深度重构模型的功率放大器行为建模方法
US20190238101A1 (en) * 2018-01-26 2019-08-01 Skyworks Solutions, Inc. Universal memory-based model for nonlinear power amplifier behaviors
CN111092602A (zh) * 2019-12-27 2020-05-01 京信通信系统（中国）有限公司 功率放大器的建模方法、装置、计算机设备和存储介质
CN113762503A (zh) * 2021-05-27 2021-12-07 腾讯云计算（北京）有限责任公司 数据处理方法、装置、设备及计算机可读存储介质
US20220200540A1 (en) * 2020-12-21 2022-06-23 Ulak Haberlesme A.S. Model trainer for digital pre-distorter of power amplifiers

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105224985A (zh) * 2015-09-28 2016-01-06 南京航空航天大学 一种基于深度重构模型的功率放大器行为建模方法
US20190238101A1 (en) * 2018-01-26 2019-08-01 Skyworks Solutions, Inc. Universal memory-based model for nonlinear power amplifier behaviors
CN111092602A (zh) * 2019-12-27 2020-05-01 京信通信系统（中国）有限公司 功率放大器的建模方法、装置、计算机设备和存储介质
US20220200540A1 (en) * 2020-12-21 2022-06-23 Ulak Haberlesme A.S. Model trainer for digital pre-distorter of power amplifiers
CN113762503A (zh) * 2021-05-27 2021-12-07 腾讯云计算（北京）有限责任公司 数据处理方法、装置、设备及计算机可读存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4546204A4 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117970004A (zh) * 2024-02-22 2024-05-03 南京莱芯科技有限公司 功率放大器的测试方法、装置、介质和电子设备
WO2025256215A1 (zh) * 2024-06-11 2025-12-18 华为技术有限公司 一种通信方法及相关装置
CN118764003A (zh) * 2024-09-09 2024-10-11 中国移动紫金(江苏)创新研究院有限公司 数据预失真处理方法、装置、设备、介质及程序产品
CN118801487A (zh) * 2024-09-14 2024-10-18 内蒙古电力（集团）有限责任公司内蒙古电力经济技术研究院分公司 一种面向新型电力系统的多类型储能资源配置方法
CN118801487B (zh) * 2024-09-14 2025-01-14 内蒙古电力（集团）有限责任公司内蒙古电力经济技术研究院分公司 一种面向新型电力系统的多类型储能资源配置方法

Also Published As

Publication number Publication date
KR20250018557A (ko) 2025-02-06
EP4546204A1 (en) 2025-04-30
EP4546204A4 (en) 2025-09-24
CN117540677A (zh) 2024-02-09
US20250315727A1 (en) 2025-10-09

Similar Documents

Publication Publication Date Title
WO2024021621A1 (zh) 功率放大器模型的获取方法、装置及功率放大器模型
WO2020098606A1 (zh) 节点分类方法、模型训练方法、装置、设备及存储介质
CN111738098A (zh) 一种车辆识别方法、装置、设备及存储介质
WO2021089013A1 (zh) 空间图卷积网络的训练方法、电子设备及存储介质
CN110598761A (zh) 一种点胶检测方法、装置及计算机可读存储介质
CN109165309B (zh) 负例训练样本采集方法、装置及模型训练方法、装置
CN112329837B (zh) 一种对抗样本检测方法、装置、电子设备及介质
CN103970879B (zh) 一种调整数据块存储位置的方法及系统
CN116305588B (zh) 一种风洞试验数据异常检测方法、电子设备及存储介质
CN115937703A (zh) 一种用于遥感图像目标检测的增强特征提取方法
WO2020082595A1 (zh) 图像分类方法、终端设备及计算机非易失性可读存储介质
CN115879012A (zh) 电力线噪声分析方法、装置、设备及存储介质
CN118469045A (zh) 多年冻土上限预测方法、装置、设备及存储介质
WO2024066143A1 (zh) 分子碰撞截面的预测方法、装置、设备及存储介质
CN111160614B (zh) 资源转移预测模型的训练方法、装置及计算设备
CN111191827A (zh) 一种数据波动趋势的预测方法及装置
CN113220883B (zh) 一种文本分类方法、装置及存储介质
CN113763193B (zh) 群体检测方法、装置、电子设备和计算机存储介质
CN111382757A (zh) 深度学习分类算法中动态调节训练样本的方法及终端设备
CN114676820A (zh) 网络模型的调优方法、装置、电子设备及存储介质
CN117743859B (zh) 工业分析模型的训练方法、使用方法及介质
CN119415845A (zh) 一种光伏功率预测方法、系统、设备和介质
CN117175546A (zh) 基于图形化建模的分布式能源配电网拓扑识别方法和介质
CN114970732A (zh) 分类模型的后验校准方法、装置、计算机设备及介质
CN121052302B (zh) 基于神经网络转换的任务推理方法、装置、设备及介质

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: 23844839

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 18875106

Country of ref document: US

ENP Entry into the national phase

Ref document number: 20257000396

Country of ref document: KR

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 1020257000396

Country of ref document: KR

WWE Wipo information: entry into national phase

Ref document number: 2023844839

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2023844839

Country of ref document: EP

Effective date: 20250121

WWP Wipo information: published in national office

Ref document number: 1020257000396

Country of ref document: KR

NENP Non-entry into the national phase

Ref country code: DE

WWP Wipo information: published in national office

Ref document number: 2023844839

Country of ref document: EP

WWP Wipo information: published in national office

Ref document number: 18875106

Country of ref document: US