WO2022062502A1 - 预测方法、装置、可读介质及电子设备 - Google Patents

预测方法、装置、可读介质及电子设备 Download PDF

Info

Publication number
WO2022062502A1
WO2022062502A1 PCT/CN2021/101314 CN2021101314W WO2022062502A1 WO 2022062502 A1 WO2022062502 A1 WO 2022062502A1 CN 2021101314 W CN2021101314 W CN 2021101314W WO 2022062502 A1 WO2022062502 A1 WO 2022062502A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
model
target device
distribution
weight
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/CN2021/101314
Other languages
English (en)
French (fr)
Inventor
杨杰
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.)
Ennew Digital Technology Co Ltd
Original Assignee
Ennew Digital Technology Co 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
Priority claimed from CN202011010347.6A external-priority patent/CN114298362A/zh
Priority claimed from CN202011010325.XA external-priority patent/CN114254790B/zh
Priority claimed from CN202011182449.6A external-priority patent/CN114429235B/zh
Application filed by Ennew Digital Technology Co Ltd filed Critical Ennew Digital Technology Co Ltd
Priority to EP21870883.2A priority Critical patent/EP4131094A4/en
Priority to JP2022564514A priority patent/JP7783832B2/ja
Publication of WO2022062502A1 publication Critical patent/WO2022062502A1/zh
Priority to US18/050,055 priority patent/US20230070276A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • 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
    • G06N3/047—Probabilistic or stochastic 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • 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
    • 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
    • G06N3/098—Distributed learning, e.g. federated learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/20—Administration of product repair or maintenance

Definitions

  • the present invention relates to the technical field of energy and the Internet of Things, and in particular, to a method and apparatus for predicting equipment failure, a method and apparatus for predicting equipment status, a readable medium, and an electronic device.
  • the industrial equipment in the intelligent manufacturing system fails, it will affect the normal operation of the intelligent manufacturing system and cause a very serious impact. Therefore, it is usually necessary to predict the equipment failure of the industrial equipment.
  • the operation data of the industrial equipment since the operation data of the industrial equipment has no label data, it is necessary to collect the historical operation data of several other industrial equipment and the label data corresponding to the historical operation data, and use the machine learning method to establish these The mapping relationship between historical operation data and equipment failure prediction is obtained, and the equipment failure prediction model is obtained, and the equipment failure prediction model of the industrial equipment is used to realize the equipment failure prediction of the industrial equipment.
  • Equipment predictive maintenance is to predict the probability of equipment failure or the remaining service life of equipment based on the characteristic information of equipment operating status.
  • the data-driven method is to use the historical operation data of the equipment, including the measurement of each sensor of the equipment, and establish the mapping relationship between these measurements and the probability of equipment failure through machine learning methods.
  • high-quality labeled data is usually required, that is, a large amount of fault label data is required; the fault data of a single device is limited, and it is necessary to combine multiple devices of the same type to obtain sufficient data.
  • the existing technical solutions can solve the problem of equipment failure prediction, but for data sharing between different devices, the existing technical solutions directly share the data of each device in the Internet of Things without considering the various devices in the Internet of Things. the need to protect data privacy.
  • Equipment condition prediction is to predict the probability of equipment failure or the remaining service life of equipment based on equipment operation data.
  • a large amount of labeled data can be obtained by obtaining the historical operation data of other equipment and the equipment status label data corresponding to the historical operation data, and then the historical operation data is established based on the machine learning method.
  • the mapping relationship between the device state label data, the device state prediction model is obtained, and the device state prediction of the device to be predicted is realized based on the device state prediction model.
  • the data distribution of historical operation data generated by different devices is different, resulting in a relatively low accuracy of the device state prediction model obtained by using the above technical solution.
  • the present invention provides a device failure prediction method, device, computer-readable storage medium and electronic device, which can migrate non-shared data to a target device and establish feature data and characteristics of the target device.
  • the relationship between equipment failures does not need to share characteristic data between equipment, ensuring data security.
  • the present invention provides a device failure prediction method, comprising: determining feature information of a target device and detection point data information corresponding to the target device; based on the feature information of the target device and the detection point data information corresponding to the target device , determine the probability distribution model of the feature data of the target device and the probability distribution model of the detection point data with unshared data; according to the probability distribution model of the feature data and the probability distribution model of the detection point data, determine the probability distribution model of the unshared data weight; establish a federated learning model according to the non-shared data, the weight of the non-shared data, and the device failure label corresponding to the non-shared data; perform device failure prediction of the target device according to the federated learning model.
  • the present invention provides an apparatus for predicting equipment failure, including: an information determination module for determining feature information of a target device and detection point data information corresponding to the target device; a probability model determination module for determining based on the target The feature information of the device and the detection point data information corresponding to the target device determine the probability distribution model of the feature data of the target device and the probability distribution model of the detection point data with non-shared data; the weight determination module is used to determine the probability distribution according to the feature data.
  • the model and the probability distribution model of the detection point data determine the weight of the non-shared data; a model building module is used for the device corresponding to the non-shared data, the weight of the non-shared data and the non-shared data according to the non-shared data
  • the fault label is used to establish a federated learning model; the prediction module is used to predict the equipment fault of the target equipment according to the federated learning model.
  • an embodiment of the present invention provides a method for predicting equipment failure.
  • the method for predicting equipment failure includes: obtaining a training data set for establishing a prediction model for the target device according to the attribute of the target device, wherein the samples in the data set are The data is shared data; the weight of each piece of sample data in the training data set is calculated; the fault prediction local model of the target equipment is obtained by using the weight training; based on the fault prediction local model and the joint learning algorithm, a joint model is established; The joint model predicts the failure of the target device.
  • an embodiment of the present invention provides an apparatus for predicting equipment faults.
  • the apparatus for predicting equipment faults includes: a data acquisition module, a weight calculation module, a local model training module, a joint model establishment module, and a fault prediction module, wherein the a data acquisition module, used for acquiring a training data set for establishing a prediction model for the target device according to the attributes of the target device, wherein the sample data in the data set is shared data; the weight calculation module is used for calculating the training data Concentrate the weight of each piece of sample data; the local model training module is used to obtain the fault prediction local model of the target device by using the weight training; the joint model building module is used to predict the fault based on the local model and joint learning. an algorithm to establish a joint model; the fault prediction module is used to predict the fault of the target device according to the joint model.
  • the present invention provides a device state prediction method, comprising: acquiring at least two feature data corresponding to a target device, at least two model training data corresponding to at least two reference devices respectively, and each of the model training data respectively The corresponding device status label, the target device and the reference device have the same device type; according to each of the feature data, determine the error weight corresponding to the model training data; according to each of the model training data, each of the model The device state labels corresponding to the training data and the error weights corresponding to each of the model training data are used for model training to determine a device state prediction model, and the device state prediction model is used for device state prediction of the target device.
  • the present invention provides a device state prediction device, comprising: an acquisition module for acquiring at least two feature data corresponding to a target device, at least two model training data corresponding to at least two reference devices respectively, and each The device status labels corresponding to the model training data respectively, the device types of the target device and the reference device are the same; the weight determination module is used to determine the error weight corresponding to the model training data according to each of the feature data; training a module for performing model training according to each of the model training data, the device state labels corresponding to each of the model training data, and the error weights corresponding to each of the model training data to determine the device state prediction model, the device state The state prediction model is used for device state prediction of the target device.
  • the present invention provides a computer-readable storage medium, comprising execution instructions, when a processor of an electronic device executes the execution instructions, the processor executes the method or the method according to any one of the first aspect.
  • the present invention provides an electronic device, comprising a processor and a memory storing execution instructions.
  • the processor executes the execution instructions stored in the memory, the processor executes the first aspect.
  • the present invention provides a device failure prediction method and device, a device state prediction method and device, a computer-readable storage medium and an electronic device.
  • the method determines the feature information of the target device and the detection point data information corresponding to the target device, and then: Based on the feature information of the target device and the detection point data information corresponding to the target device, determine the probability distribution model of the feature data of the target device and the probability distribution model of the detection point data with non-shared data, and then according to the feature data probability distribution model and detection point data Then, according to the weight of the non-shared data, the weight of the non-shared data and the equipment fault label corresponding to the non-shared data, a federated learning model is established, and finally, according to the federated learning model, the equipment of the target device is failure prediction.
  • the technical scheme provided by the present invention establishes the relationship between the feature data of the target device and the device failure by migrating the non-shared data to the target device, without sharing the feature data between devices and ensuring data security
  • FIG. 1 is a schematic flowchart of a method for predicting equipment failure according to an embodiment of the present invention
  • FIG. 2 is a schematic flowchart of another equipment failure prediction method provided by an embodiment of the present invention.
  • FIG. 3 is a schematic flowchart of a method for predicting equipment failure according to an embodiment of the present invention.
  • FIG. 4 is a schematic flowchart of another equipment failure prediction method provided by an embodiment of the present invention.
  • FIG. 5 is a schematic flowchart of a device state prediction method according to an embodiment of the present invention.
  • FIG. 6 is a schematic flowchart of another device state prediction method provided by an embodiment of the present invention.
  • FIG. 7 is a schematic structural diagram of an apparatus for predicting equipment failure according to an embodiment of the present invention.
  • FIG. 8 is a schematic structural diagram of an apparatus for predicting equipment failure according to an embodiment of the present invention.
  • FIG. 9 is a schematic structural diagram of an apparatus for predicting equipment state according to an embodiment of the present invention.
  • FIG. 10 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
  • FIG. 1 is a schematic flowchart of a method for predicting equipment failure according to an embodiment of the present invention. As shown in FIG. 1 , an embodiment of the present invention provides a method for predicting equipment faults, including the following steps:
  • Step 101 Determine feature information of the target device and detection point data information corresponding to the target device.
  • the target device is any industrial device of an intelligent manufacturing system, which is not specifically limited in this embodiment of the present invention, and preferably, the target device is an energy device, such as a gas boiler, an internal combustion engine, a steam turbine, a cogeneration device, a photovoltaic device equipment, etc.
  • an energy device such as a gas boiler, an internal combustion engine, a steam turbine, a cogeneration device, a photovoltaic device equipment, etc.
  • the feature information of the target device refers to attributes or functions of the feature data of the target device.
  • the detection point data information refers to attributes or functions of the detection point data.
  • the detection point data includes non-shared data.
  • Step 102 based on the feature information of the target device and the detection point data information corresponding to the target device, determine a probability distribution model of the feature data of the target device and a probability distribution model of the detection point data with non-shared data.
  • each feature data includes several features corresponding to the feature values of the target device, wherein several features are the influencing factors of the device failure, which need to be determined in combination with the actual scene, for example, If the target device is a gas boiler, several features include but are not limited to gas flow, gas temperature, exhaust gas temperature, gas flow, gas pressure, on-off state, flue gas humidity, and flue gas pressure.
  • the feature data is subject to continuous probability distribution.
  • the feature data probability distribution model may be a normal distribution model or an exponential distribution model, which is not limited in this embodiment of the present invention.
  • the feature data probability distribution model is A mixed Gaussian model composed of multiple normal distribution models, wherein the normal distribution model is also called a Gaussian distribution model.
  • the probability distribution model of the detection point data is the same as the probability distribution model of the feature data.
  • it can be a Gaussian mixture model, and of course, it can also be other probability distribution models, which need to be determined in combination with the actual situation, which is not specifically limited here. .
  • the feature data probability distribution model of the target device and the probability distribution model of the detection point data with non-shared data are determined based on the feature information of the target device and the detection point data information corresponding to the target device.
  • the feature data probability distribution model includes: based on the feature information of the target device and the detection point data information corresponding to the target device, determining the feature data of the target device and the detection point data with non-shared data; calculating according to the feature data parameter model
  • the data distribution of the characteristic data, and the characteristic data parameter model of the determined parameters is determined as the characteristic data probability distribution model; the data distribution of the detection point data is calculated according to the detection point data parameter model, and the determined parameters are determined.
  • the detection point data parameter model is determined as a probability distribution model of the detection point data.
  • the feature data parameter model may be a Gaussian model or a mixed Gaussian model, preferably a mixed Gaussian model.
  • the detection point data parameter model is the same as the feature data parameter model, which will not be described in detail in this embodiment of the present invention.
  • the feature data probability distribution model is to determine the model parameters by substituting each feature data into the mixture Gaussian model for calculation, wherein the calculation method may be the EM algorithm, and the mixture Gaussian model with the determined model parameters is determined as the feature data probability distribution. Model.
  • the probability distribution model of the detection point data is to determine the model parameters by substituting multiple detection point data into the mixture Gaussian model for calculation.
  • the calculation method can be the EM algorithm, and the mixture Gaussian model with the determined model parameters is determined as the detection point A probability distribution model for the data.
  • Step 103 Determine the weight of the non-shared data according to the probability distribution model of the feature data and the probability distribution model of the detection point data.
  • the non-shared data of the detection point is migrated to the target device through the weight of the non-shared data, in other words, the data association between the detection point and the target device is established, and the detection point will not directly obtain the target device.
  • the characteristic data of the device that is, there is no data sharing between the target device and the detection point, thus ensuring data security.
  • the determining the weight of the non-shared data according to the feature data probability distribution model and the probability distribution model of the detection point data includes: determining, according to the feature data probability distribution model, the The target device distribution probability of the non-shared data; according to the probability distribution model of the detection point data, the detection point distribution probability of the non-shared data is determined; the target device distribution probability of the non-shared data and the non-shared data The ratio of the detection point distribution probability is determined as the weight of the non-shared data.
  • the weight of the non-shared data is determined by the ratio of the target device distribution probability of the non-shared data to the detection point distribution probability of the non-shared data, so that the non-shared data on the detection point is migrated to the target device, and at the same time There is no need to share the characteristic data of the target device, ensuring data security.
  • the non-shared data is substituted into the feature data probability distribution model, and the value output by the feature data probability distribution model is the target device distribution probability of the non-shared data.
  • the non-shared data is substituted into the probability distribution model of the detection point data, and the value output by the probability distribution model of the detection point data is the detection point distribution probability of the shared data.
  • Step 104 Establish a federated learning model according to the non-shared data, the weight of the non-shared data, and the device fault label corresponding to the non-shared data.
  • a federated learning model is established based on the non-shared data, the weight of the non-shared data, and the device fault label corresponding to the non-shared data, so as to realize the migration of the non-shared data to the target device without sharing the feature data of the target device, thereby Data security is ensured.
  • the feature data and detection point data of the target device are distributed in different detection points in the Internet of Things, and sharing data for model training will cause data security problems.
  • the weight of the non-shared data and the device fault label corresponding to the non-shared data realize the migration of the non-shared data to the target device, so that there is no data sharing between the detection points, and the data security problem caused by the direct data sharing is avoided.
  • the detection point is a node capable of data processing and data interaction, including but not limited to any one or more of an edge server, an edge gateway, and an edge controller.
  • the equipment fault label may be the feature information of the operating state of the equipment to predict the probability of equipment failure.
  • the equipment fault label may be the type of fault or the degree of fault, which is not limited in this embodiment of the present invention, and needs to be combined with actual requirements.
  • this embodiment of the present invention does not intend to limit the method for obtaining the device fault label corresponding to each non-shared data, which may also be manual labeling, rule labeling, or cluster labeling.
  • the labeling methods are all in the prior art, and will not be repeated here.
  • the device failure prediction method provided by the embodiment of the present invention can also be used to predict the remaining service life of the device. Specifically, the device failure label can be replaced by the remaining service life of the device.
  • the establishing a federated learning model according to the non-shared data, the weight of the non-shared data, and the device fault label corresponding to the non-shared data includes: according to the non-shared data, the The weight of the non-shared data and the equipment fault label corresponding to the non-shared data are used to determine a local model of the detection point; a federated learning model is established according to at least two of the local models of the detection point.
  • model training is performed based on non-shared data, weights corresponding to non-shared data, and equipment fault labels to determine a local model of the detection point, and then the local models of multiple detection points are fused to establish a federated learning model.
  • the non-shared data and the weights of the non-shared data can realize the migration of the non-shared data of the detection point to the target device.
  • the model parameters are adjusted by the respective weights of the multiple non-shared data, so that the adjusted model parameters can be It reflects the relationship between the feature data of the target device and the device failure, and does not involve the sharing of the detection point data and the feature data of the target device, thereby ensuring data security.
  • establishing a federated learning model according to at least two local models of detection points includes: based on a joint learning algorithm, repeatedly iterating on at least two local models of detection points, and based on at least two local models of detection points.
  • a federated learning model is established based on the iterative good local model of the two detection points.
  • the joint learning algorithm refers to sending the local model parameters of the detection point in the local model of each detection point to the detection point of the target device, and the detection point of the target device averages or weights the local model parameters of each detection point to obtain federated learning.
  • Model parameters based on the federated learning parameters, the local model update iteration of the detection point is performed.
  • the detection point of the target device fuses the model parameters of the detection point local model from each detection point to obtain the federated learning model parameters, Distribute the fused federated learning model parameters to the detection point local models of each detection point, and then perform local training according to the unshared data of each detection point and the fused joint model parameters to obtain the updated model parameters.
  • the model parameters are sent to the detection point of the target device, and the iteration is repeated according to the above method until the number of iterations meets the preset number of times, or the model error of the local model of the iterative detection point meets the preset value.
  • the local model parameters of the iterative detection points are fused to obtain the federated learning model parameters, and then the federated learning model is obtained.
  • the manner of fusion of the model parameters of the local models of each detection point may be average or weighted average, which is not limited in this embodiment of the present invention.
  • the local model of the detection point may be a neural network model or a regression model, which needs to be determined according to actual requirements.
  • the detection point local model can be iterated in the following ways:
  • A1. Perform model training according to a plurality of non-shared data, the corresponding equipment fault label of each non-shared data, and the corresponding weight of each non-shared data to determine the local model of the detection point;
  • A2 Determine whether the model error of the local model of the detection point satisfies the iterative condition. If so, determine the local model of the detection point as the detection point sent by the final model to the target device, if not, execute A3;
  • A3. Send the model parameters of the local model of the detection point to the detection point of the target device;
  • A4 Receive the fused model parameters sent by the detection point of the target device, and analyze the fused model according to the multiple unshared data, the equipment fault labels corresponding to the multiple unshared data, and the weights corresponding to the multiple unshared data. The parameters are adjusted to determine the adjusted model parameters, and the adjusted model parameters are replaced with the model parameters of the local model of the detection point, and A2 is executed.
  • Step S105 establishing a federated learning model according to the local models of at least two of the detection points.
  • the current feature data of the target device is collected, and the current feature data is substituted into the federated learning model to determine whether the target device is faulty.
  • A determines the probability distribution model of the non-shared data according to the mixed Gaussian model, and receives the probability distribution model of the feature data sent by D, so as to determine the corresponding weight of the non-shared data, B and C according to
  • the above-mentioned A is similar to the processing process, and the weight corresponding to each non-shared data in B and C is obtained.
  • A, B, and C are respectively based on each non-shared data, the device fault label corresponding to the non-shared data, and each non-shared data.
  • D performs federated learning on the local models of detection points sent by A, B, and C to obtain the federated learning model of the target device.
  • the embodiments of the present invention have at least the following effective effects: the weight of the non-shared data is determined by the probability distribution model of the feature data and the probability distribution model of the non-shared data, so as to obtain the local model of the detection point of the target device, and the non-shared data is established.
  • the data association between the detection point of the data and the target device, the non-shared data is migrated to the target device, and at the same time, a federated learning model is established based on the non-shared data, and the federated learning model is used to realize the device failure prediction of the target device. There is no need to share device characteristic data with the target device, thus ensuring data security.
  • FIG. 1 shows only a basic embodiment of the method of the present invention, and other preferred embodiments of the method can also be obtained by performing certain optimizations and expansions on the basis.
  • FIG. 2 is a schematic flowchart of another equipment fault prediction method provided by an embodiment of the present invention. This embodiment is further described in combination with specific application scenarios on the basis of the foregoing embodiments. As shown in Figure 2, the equipment failure prediction method may specifically include the following steps:
  • Step 201 Determine feature information of the target device and detection point data information corresponding to the target device.
  • the detection point corresponding to the target device in the Internet of Things is n, and there are N non-shared data detection points, and the corresponding detection points in the Internet of Things are n1, n2, ..., nN, n1, n2 respectively. , ..., nN can obtain non-shared data respectively, and n can obtain all characteristic data of the target device. It should be noted that the process of obtaining the local model of the detection point by n1, n2, ..., nN is similar, and only the processing process of n1 will be described below.
  • Step 202 based on the feature information of the target device and the mixture Gaussian model, calculate the data distribution of the feature data corresponding to the feature information, and determine the mixture Gaussian model with the determined model parameters as the feature data probability distribution model.
  • each feature data is substituted into the mixture Gaussian model to calculate each feature data, thereby obtaining each feature data probability distribution model, and each feature data probability distribution model is sent to n1, n2, ..., nN respectively.
  • n1 The process of obtaining the probability distribution model of the non-shared data by n1, n2, ..., nN is similar, and the processing process of n1 is described below.
  • Step 203 based on the detection point data information corresponding to the target device and the mixture Gaussian model, calculate the data distribution of the detection point data with non-shared data corresponding to the detection point data information, and determine the mixture Gaussian model of the model parameters as all the parameters. Describe the probability distribution model for unshared data.
  • n1 acquires multiple non-shared data, and substitutes the non-shared data into the mixture Gaussian model for calculation to obtain the probability distribution model of the non-shared data of n1.
  • Step 204 Determine the target device distribution probability of the non-shared data according to the feature data probability distribution model, and determine the detection point distribution probability of the non-shared data according to the probability distribution model of the detection point data.
  • n1 receives the probability distribution model corresponding to the feature data, and substitutes the obtained non-shared data into the feature data probability distribution model for calculation to obtain the target device distribution probability of the non-shared data.
  • n1 substitutes the non-shared data into the probability distribution model of the non-shared data for calculation, and obtains the detection point distribution probability of the non-shared data.
  • Step 205 Determine the ratio of the target device distribution probability of the non-shared data to the detection point distribution probability of the non-shared data as the weight of the non-shared data.
  • the weight of each non-shared data is obtained, In this way, the data link between the target device and the detection point that does not share data can be established.
  • Step 206 Determine the local model of the detection point according to the non-shared data, the weight of the non-shared data, and the device fault label corresponding to the non-shared data.
  • n1 performs model training according to the non-shared data, the weight of the non-shared data, and the equipment fault label corresponding to the non-shared data, and determines the local model of the detection point of the target device.
  • n2, ..., nN respectively determine the local model of the detection point of the target device according to the method similar to n1.
  • Step 207 based on the joint learning algorithm, repeatedly iterate the detection point local models of at least two of the detection points, and establish a federated learning model based on the iterated local detection point models of the at least two detection points.
  • n and n1, n2, ..., nN using a joint learning algorithm, the local models of detection points trained respectively by n1, n2, ..., nN are modeled after model iteration and sent to n, n pairs n1, n2, ..., The iterative local models of detection points obtained by nN training are averaged, and the federated learning model of the target device is established.
  • Step 208 performing the equipment failure prediction of the target equipment on the federated learning model.
  • the equipment failure prediction of the target equipment is carried out.
  • the beneficial effect of this embodiment is that the weight of the non-shared data is obtained through the ratio of the probability distribution model of the non-shared data and the probability distribution model of the characteristic data, and the detection point of the non-shared data and the target device are established.
  • the data association between the two migrate the non-shared data to the target device, and realize the establishment of the federated learning model of the target device through the detection points of the target device and the non-shared data in the Internet of Things, so as to realize the detection point of the target device and the non-shared data.
  • FIG. 3 is a schematic flowchart of a method for predicting equipment failure according to an embodiment of the present invention. As shown in Figure 3, the equipment failure prediction method includes:
  • Step 301 according to the attribute of the target device, obtain a training data set for establishing a prediction model for the target device, wherein the sample data in the data set is shared data;
  • Step 302 calculating the weight of each piece of sample data in the training data set
  • Step 303 using the weight training to obtain a fault prediction local model of the target device
  • Step 304 establishing a joint model based on the fault prediction local model and the joint learning algorithm
  • Step 305 Perform fault prediction on the target device according to the joint model.
  • the sample data in the training data set includes feature data of the target device, feature data of the sample device, and fault data of the sample device; the sample device is a device related to or similar to the target device.
  • the fault prediction local model and the joint model are the relationship between the characteristic data of the equipment and the fault of the equipment.
  • the data of different devices is distributed in different nodes of the Internet of Things, and the shared data training model will cause data security problems.
  • This embodiment adopts a sample migration method based on the joint learning method, which is used for predictive maintenance of equipment, and can be transferred on multiple devices.
  • the collected data is learned jointly and migrated to the target device to train the predictive maintenance model for the target device, realizing multi-party joint learning, ensuring that the data of all parties is not local, and avoiding the data security problem caused by the direct sharing of data .
  • the calculating the weight of each piece of sample data in the training data set includes, for each of the sample devices: distinguishing the feature data of the sample device from the feature data of the target device ; Train a classification model according to the distinguished feature data; calculate the weight of each feature data of the sample device according to the trained classification model.
  • a classification model based on joint learning can be adopted, such as the XGBoost model based on joint learning in this embodiment.
  • the calculating the weight of each piece of sample data in the training data set includes, for each of the sample devices: marking the characteristic data of the sample device as first data, and marking the feature data of the sample device as first data, The feature data of the prediction device is marked as second data; according to the first data and the second data, a classification model is trained, and the classification model is a classification model based on joint learning; according to the trained classification model, the The weight of each feature data of the sample device, the calculation formula of the weight is:
  • ⁇ i is the weight of the i-th piece of data in the first data
  • P 1i is the probability that the i-th piece of data belongs to the sample device
  • P 2i is the probability that the i-th piece of data belongs to the target device probability.
  • the obtaining a local model of fault prediction of the target device by using the weight training includes: according to the feature data of each sample device, the weight of each feature data of the sample device, and each device For the fault data of the sample device, a neural network is used to train on the training data set with weights of the sample device, respectively, to obtain a local model of failure prediction of the target device.
  • the establishing a joint model based on the fault prediction local model and the joint learning algorithm includes: repeating iteratively using the joint learning algorithm according to the fault prediction local model, to obtain the joint model on the target device on the training dataset.
  • the target device and each of the sample devices are edge nodes in the Internet of Things
  • the feature data of the target device is not exposed to other sample devices
  • the features of each sample device are Data and failure data are not exposed to other said sample devices and said target device.
  • FIG. 4 is a schematic flowchart of another equipment fault prediction method provided by an embodiment of the present invention. As shown in Figure 4, the equipment failure prediction method includes:
  • Step 401 acquiring sensor characteristic data of each sample device and corresponding fault data; collecting sensor characteristic data of the target device;
  • Step 402 label the feature data of the target device and the sample device respectively to obtain the label data
  • Step 403 based on the label data of step 402, adopt the method of joint learning to train the classifier (two classification);
  • Step 404 calculate the weight of the characteristic data of the sample device
  • Step 405 repeating the above steps to calculate the weight of the characteristic data of each sample device
  • Step 406 based on the data of each sample device, establish the relationship between the sensor feature data and the failure of the device by means of joint learning;
  • step 407 the model in step 406 is used to predict the failure probability of the target equipment.
  • the goal is to use the feature data of devices A, B, and C, the fault data, and the feature data of device D to learn a probability model for predicting the failure of device D, which is used for device D's failure prediction.
  • the equipment failure prediction method includes:
  • the training data set includes all characteristic data and fault data of equipment A, B, and C, as well as characteristic data of equipment D;
  • train the classifier (two classification) (in this example, the XGBoost model based on joint learning is used, in practice, XGBoost is not limited, and any probability classification model based on joint learning can be used) ;
  • the joint model for failure prediction of the device D is also the failure prediction model; in this embodiment, the number of repeated iterations is determined by the preset precision of the model and the preset number of iterations.
  • the equipment failure data mentioned in the present invention can also be replaced with the remaining service life of the equipment, so as to predict the remaining service life of the equipment.
  • FIG. 5 is a schematic flowchart of a device state prediction method according to an embodiment of the present invention. As shown in Figure 5, the equipment state prediction method includes the following steps:
  • Step 501 Acquire at least two feature data corresponding to a target device, at least two model training data corresponding to at least two reference devices, and device status labels corresponding to each of the model training data, the target device and the reference device respectively.
  • Devices are of the same kind of device.
  • the operation data of the reference device may be acquired according to the sensors installed on the reference device connected to the network, and the plurality of model training data corresponding to the reference device may be determined according to the operation data of the reference device. It should be noted that the acquisition of model training data based on data collected by sensors is in the prior art, which will not be repeated in this embodiment.
  • the operation data of the target device can be acquired according to the sensors installed in the target device connected to the network, and the plurality of characteristic data corresponding to the target device can be determined according to the operation data of the target device. It should be noted that the acquisition of characteristic data based on data collected by sensors is in the prior art, which is not described in detail in this embodiment.
  • the model training data includes several features, and each feature corresponds to one or more feature values of the reference device, preferably one, wherein several features are influencing factors that have an impact on the state of the device.
  • Several features corresponding to the model training data and feature data are the same, and the feature values corresponding to the features may be different.
  • the model training data corresponds to a device state label respectively. It should be noted that the embodiments of the present invention do not intend to limit the number of reference devices, which need to be determined in combination with actual scenarios.
  • the device status tag includes, but is not limited to, device failure information or the remaining useful life of the device. It should be noted that this embodiment of the present invention does not intend to limit the method for obtaining the device status label corresponding to each model training data, which may also be manual labeling, rule labeling, or cluster labeling. The labeling methods are all in the prior art, and will not be repeated here.
  • both the reference device and the target device are gas boilers.
  • the feature data of the target device has no device state label, and the model training cannot be completed with its own data. Therefore, it is necessary to obtain the corresponding device state prediction model with the help of the model training data of the reference device.
  • Step 502 Determine the error weight corresponding to the model training data according to each of the characteristic data.
  • the error weight corresponding to the model training data can be determined by the following method:
  • the replacement data Preferably, the data in the multiple feature data is the same as the model training data. If the same data as the model training data does not exist in the multiple feature data, the replacement data can be the most similar data to the model training data among the multiple feature data. , as a possible situation, when there are two or more feature data of the training data of the replaceable model in the plurality of feature data, at this time, one feature data of the training data of the replaceable model can be randomly selected as the replacement data.
  • first distribution information of the model training data in each model training data corresponding to the reference device to which it belongs is determined.
  • the first distribution information indicates the distribution of all model training data corresponding to the reference device to which the model training data belongs.
  • the first distribution information may be a distribution probability or an occurrence frequency, which is not specifically limited here.
  • the second distribution information of the replacement data in each feature data is determined.
  • the second distribution information indicates the distribution of all feature data of the replacement data in the target device, and the second distribution information may be a distribution probability or an occurrence frequency, which is not specifically limited here.
  • the contents included in the first distribution information and the second distribution information should be unified.
  • the first distribution information and the second distribution information may both include distribution probabilities, or both may include occurrence frequencies, which will not be detailed here. limited.
  • this embodiment does not intend to limit the method for determining the distribution information, any method used to determine the data distribution in the prior art is sufficient, for example, it may be a parameter estimation method or a non-parametric estimation method, It can also be a semi-parametric estimation method.
  • the first distribution information is the first distribution probability
  • the second distribution information is the second distribution probability.
  • the first distribution probability and the second distribution probability may be determined in the following manner:
  • Non-parametric estimation methods are not limited by the overall distribution, do not assume the specific form of the overall distribution, try to obtain the required information from the data or the sample itself, and obtain the structure of the distribution through estimation.
  • the dynamic structure plays an extremely important role, especially when the data distribution is very irregular, the non-parametric estimation method is more accurate than the parameter estimation, so in this embodiment, the non-parametric estimation method is used to determine the distribution probability.
  • the probability density function of each model training data corresponding to the reference device is determined by the first non-parametric estimation method, based on The probability density distribution function thus determines the first distribution probability of each model training data corresponding to the reference device.
  • the probability density function of each feature data corresponding to the target device is determined by the second nonparametric estimation method, and the second distribution probability of each feature data corresponding to the target device is determined based on the probability density distribution function.
  • the accuracy of the estimated data distribution may be reduced. Therefore, it is necessary to combine the data volume of each reference device and the target device and Data distribution, select an appropriate nonparametric estimation method to estimate the data distribution.
  • the difference between the data volume of each model training data corresponding to each reference device and the data volume of each feature data corresponding to the target device is small, the first non-parameter estimation method corresponding to each reference device and the first non-parameter estimation method corresponding to the target device
  • the second nonparametric estimation methods are the same. As a possible situation, the first nonparametric estimation method and the second nonparametric estimation method are both kernel density estimation methods.
  • the error weight corresponding to the model training data may be specifically determined in the following manner:
  • a distribution probability ratio between the second distribution probability of the replacement data corresponding to the model training data and the first distribution probability of the model training data is determined, and the distribution probability ratio is determined as the error weight corresponding to the model training data.
  • the error weight corresponding to the model training data may be specifically determined in the following manner:
  • the ratio of the second occurrence frequency of the replacement data corresponding to the model training data to the first occurrence frequency of the model training data is determined, and the occurrence frequency ratio is determined as the error weight corresponding to the model training data.
  • Step 503 Perform model training according to each of the model training data, the device state labels corresponding to each of the model training data, and the error weights corresponding to each of the model training data, to determine a device state prediction model, the device state
  • the prediction model is used for device state prediction of the target device.
  • the device state prediction model may be a neural network model or a regression model, which is not limited here, and needs to be determined in combination with the actual situation.
  • the device state prediction method provided in this embodiment is applied to an edge server or a local server corresponding to the target device, thereby reducing the computing pressure of the cloud server.
  • the equipment state prediction model uses multiple model training data and equipment state label data corresponding to each model training data as training data, and adjusts the model parameters of the equipment state prediction model through the error weights corresponding to each model training data, respectively, Therefore, the adjusted model parameters can reflect the relationship between the characteristic data of the target equipment and the equipment state, and the model accuracy of the equipment state prediction model is ensured.
  • the device state prediction model can be determined by:
  • A2 Determine whether the number of iterations is satisfied or whether the second error corresponding to each model training data satisfies the preset condition. If so, determine the to-be-predicted model as the equipment state prediction model, and if not, execute A3;
  • whether the second error corresponding to each model training data satisfies the preset condition may be whether the mean value of the second error corresponding to each model training data is smaller than a threshold, or it may be the standard of the second error corresponding to each model training data Whether the variance is less than the threshold is not specifically limited here.
  • the predicted result can be understood as equipment status information
  • the first error is the degree of proximity between the predicted result and the equipment status label.
  • the predicted result can be a fault or no fault.
  • the equipment status label Also in failure the first error may be 0, otherwise 1.
  • the beneficial effect of this embodiment is: by determining the error weights of the training data of the models corresponding to the reference devices with respect to the target device, respectively, the model training data corresponding to the reference devices can be divided into The training data is migrated to the target device, and subsequently, model training is performed using multiple model training data corresponding to multiple reference devices, the device status label corresponding to each model training data, and the weight error corresponding to each model training data.
  • the obtained device state prediction model has relatively high accuracy because the error weight of the model training data for the target device is comprehensively considered.
  • a more accurate device state of the target device can be obtained.
  • Fig. 5 shows only a basic embodiment of the method of the present invention, and other preferred embodiments of the method can also be obtained by performing certain optimization and expansion on the basis.
  • FIG. 6 is a schematic flowchart of another device state prediction method provided by an embodiment of the present invention. Based on the foregoing embodiments, this embodiment is described in more detail in combination with application scenarios. As shown in Figure 6, the device state prediction method specifically includes the following steps:
  • Step 601 Acquire at least two feature data corresponding to the target device, at least two model training data corresponding to at least two reference devices, and device status labels corresponding to each of the model training data, the target device and the reference device.
  • Devices are of the same kind of device.
  • Boiler A, boiler B, and boiler C respectively have N model training data, and each model training data corresponds to a fault label.
  • the fault label can be fault or normal
  • boiler D corresponds to n feature data
  • the feature data and model training data respectively correspond to several features that are the same, but the feature values corresponding to each feature may be different.
  • Step 602 Determine the first distribution probability of the model training data in each of the model training data corresponding to the reference device to which it belongs, based on the first non-parametric estimation method corresponding to the model training data in the reference device to which it belongs. .
  • boiler B and boiler C are all kernel density estimation methods
  • the probability density distribution function of boiler A corresponding to N model training data for each model training data in boiler A, based on the probability density distribution function of boiler A corresponding to N model training data, determine the first distribution probability P1 of the model training data in all model training data in boiler A.
  • the calculation method of the first distribution probability P1 of each model training data corresponding to boiler B and boiler C, respectively, is similar to that of boiler A, and will not be repeated here.
  • Step 603 for each of the model training data, determine the replacement data corresponding to the model training data from each of the feature data; based on the second nonparametric estimation method, determine that the replacement data is in each of the feature data.
  • the second distribution probability in .
  • Step 604 For each of the model training data, determine the distribution probability ratio between the second distribution probability of the replacement data corresponding to the model training data and the first distribution probability of the model training data; determine the distribution probability ratio In order to determine the error weight corresponding to the model training data.
  • Step 605 Perform model training according to each of the model training data, the device state labels corresponding to each of the model training data, and the error weights corresponding to each of the model training data to determine a device state prediction model, the device state
  • the prediction model is used for device state prediction of the target device.
  • the equipment state prediction model is trained with N model training data corresponding to boiler A, boiler B, and boiler C respectively, and the fault labels corresponding to each model training data, respectively, based on the error weights corresponding to each model training data. W adjusts the model parameters, and finally obtains the equipment state prediction model.
  • FIG. 7 is a schematic structural diagram of an apparatus for predicting equipment failure according to an embodiment of the present invention.
  • the equipment state failure prediction device includes:
  • An information determination module 701 configured to determine feature information of the target device and detection point data information corresponding to the target device;
  • a probability model determination module 702 configured to determine the feature data probability distribution model of the target device and the probability distribution model of the detection point data with non-shared data based on the feature information of the target device and the detection point data information corresponding to the target device;
  • a weight determination module 703, configured to determine the weight of the non-shared data according to the feature data probability distribution model and the probability distribution model of the detection point data;
  • a model establishment module 704 configured to establish a federated learning model according to the non-shared data, the weight of the non-shared data, and the device fault label corresponding to the non-shared data;
  • the prediction module 705 is configured to perform device failure prediction of the target device according to the federated learning model.
  • the probability model determination module 702 includes: a data determination unit, a first probability distribution model determination unit, and a second probability distribution model determination unit; wherein, the data determination unit is configured to determine based on the target The feature information of the device and the detection point data information corresponding to the target device are used to determine the feature data of the target device and the detection point data with non-shared data; the first probability distribution model determining unit is used for the feature information of the target device and the target The detection point data information corresponding to the device determines the characteristic data of the target device and the detection point data with non-shared data; the second probability distribution model determination unit is used for calculating the data distribution of the detection point data according to the detection point data parameter model, The detection point data parameter model whose parameters have been determined is determined as the probability distribution model of the detection point data.
  • the feature data parameter model and/or the detection point data parameter model includes a Gaussian mixture model.
  • the weight determination module 703 includes: a first distribution probability determination unit, a second distribution probability unit, and a weight determination unit; wherein the first distribution probability determination unit is configured to a feature data probability distribution model, for determining the target device distribution probability of the non-shared data; the second distribution probability unit for determining the detection point distribution probability of the non-shared data according to the probability distribution model of the detection point data ; the weight determination unit is configured to determine the ratio of the target device distribution probability of the non-shared data to the detection point distribution probability of the non-shared data as the weight of the non-shared data.
  • the model establishment module 704 includes: a local model determination unit and a model establishment unit; the local model determination unit is configured to determine the non-shared data according to the weight and the weight of the non-shared data.
  • the equipment fault label corresponding to the non-shared data determines a local model of the detection point; the model building unit is configured to establish a federated learning model according to the local models of the detection point of at least two of the detection points.
  • the model building unit is configured to repeatedly iterate the local models of the detection points of at least two of the detection points based on a joint learning algorithm, and based on the iterative results of the at least two detection points
  • the detection point local model is established to establish a federated learning model.
  • the detection point local model includes a neural network model or a regression model.
  • FIG. 8 is a schematic structural diagram of an apparatus for predicting equipment failure according to an embodiment of the present invention.
  • the equipment failure prediction device includes: a data acquisition module 801, a weight calculation module 802, a local model training module 803, a joint model establishment module 804, and a failure prediction module 805, wherein,
  • the data acquisition module 801 is configured to acquire a training data set for establishing a prediction model for the target device according to the attribute of the target device, wherein the sample data in the data set is shared data;
  • the weight calculation module 802 is used to calculate the weight of each piece of sample data in the training data set
  • the local model training module 803 is used to obtain a fault prediction local model of the target device by using the weight training;
  • the joint model establishment module 804 is configured to establish a joint model based on the fault prediction local model and the joint learning algorithm
  • the fault prediction module 805 is configured to perform fault prediction on the target device according to the joint model.
  • the sample data in the training data set includes feature data of the target device, feature data of the sample device, and fault data of the sample device; the sample device is related to or similar to the target device device of.
  • the weight calculation module includes a data labeling unit, a data classification unit, and a data calculation unit, wherein, for each of the sample devices, the data labeling unit is used for the data acquisition module
  • the acquired feature data of the sample device is distinguished from the feature data of the target device;
  • the data classification unit is configured to train a classification model according to the feature data differentiated by the data labeling unit;
  • the data calculation unit According to the classification model trained by the data classification unit, the weight of each characteristic data of the sample device is calculated.
  • the data marking unit is specifically configured to mark the characteristic data of the sample device acquired by the data acquisition module as first data, and label the data The characteristic data of the target device acquired by the acquisition module is marked as second data;
  • the data classification unit is specifically configured to train a classification model according to the first data and the second data marked by the data marking unit,
  • the classification model is a classification model based on joint learning;
  • the data calculation unit is specifically configured to calculate the weight of each feature data of the sample device according to the classification model trained by the data classification unit, and the calculation of the weight
  • the formula is:
  • ⁇ i is the weight of the i-th piece of data in the first data
  • P 1i is the probability that the i-th piece of data belongs to the sample device
  • P 2i is the probability that the i-th piece of data belongs to the target device probability.
  • the local model training module is specifically configured to use the feature data of each sample device, the weight of each feature data of the sample device, and the fault data of each sample device
  • the neural network is respectively trained on the training data sets with weights for the sample devices to obtain a local model of failure prediction of the target device.
  • the joint model building module is specifically configured to, according to the fault prediction local model, use a joint learning algorithm iteratively to obtain the information about the target device of the sample device on the training data set. joint model.
  • the target device and each of the sample devices are edge nodes in the Internet of Things
  • the feature data of the target device is not exposed to other sample devices
  • the features of each sample device are Data and failure data are not exposed to other said sample devices and said target device.
  • FIG. 9 is a schematic structural diagram of an apparatus for predicting equipment state according to an embodiment of the present invention.
  • the device state prediction device includes:
  • the acquisition module 901 is configured to acquire at least two feature data corresponding to the target device, at least two model training data corresponding to at least two reference devices respectively, and device status labels corresponding to each of the model training data, the target device and The reference devices are of the same device type;
  • a weight determination module 902 configured to determine an error weight corresponding to the model training data according to each of the characteristic data
  • the training module 903 is configured to perform model training according to each of the model training data, the device state labels corresponding to each of the model training data, and the error weights corresponding to each of the model training data, so as to determine the device state prediction model.
  • the device state prediction model is used for device state prediction of the target device.
  • the weight determination module 902 includes: a replacement data determination unit, a distribution information determination unit, and an importance degree determination unit; wherein the replacement data determination unit is configured to determine from each of the feature data Replacement data corresponding to the model training data; the distribution information determining unit is configured to determine the first distribution information and the replacement data of the model training data in each of the model training data corresponding to the reference device to which it belongs the second distribution information in each of the feature data; the importance determination unit is configured to determine, according to the first distribution information of the model training data and the second distribution information of the replacement data corresponding to the model training data, The error weight corresponding to the model training data.
  • the distribution information determination unit includes: a first information determination subunit and a second information determination subunit; wherein the first information determination subunit is configured to The first non-parameter estimation method corresponding to the reference device to which it belongs, determines the first distribution probability of the model training data in each of the model training data corresponding to the reference device to which it belongs, and determines the first distribution probability as the first distribution probability.
  • the second information determination subunit is configured to determine the second distribution probability of the replacement data in each of the feature data based on the second non-parametric estimation method, and determine the second distribution probability as Second distribution information.
  • both the first nonparametric estimation method and the second nonparametric estimation method are kernel density estimation algorithms.
  • the weight determination module 302 is configured to, for each of the model training data, determine a second distribution probability of the replacement data corresponding to the model training data and a first distribution probability of the model training data The distribution probability ratio is determined as the error weight corresponding to the model training data.
  • the method is applied to an edge server or a local server corresponding to the target device.
  • the training module 903 includes: a training unit, a judging unit, and an adjusting unit; wherein, the training unit is configured to substitute the model training data into the prediction result in the to-be-predicted model and the The device status label corresponding to the model training data, determining the first error corresponding to the model training data, and determining the second error corresponding to the model training data according to the first error corresponding to the model training data and the error weight;
  • the judging unit is used for judging whether the number of iterations is satisfied or whether the second error corresponding to each of the model training data satisfies a preset condition, if so, the to-be-predicted model is determined as the equipment state prediction model, if not, then trigger the adjustment unit; the adjustment unit is configured to adjust the model parameters in the to-be-predicted model according to the second error corresponding to each of the model training data, so as to determine the adjusted model parameters, and The model parameters in the to-be-predicted model are replaced with the adjusted model parameters, and
  • FIG. 10 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention.
  • the electronic device includes a processor 1001 , a memory 1002 storing execution instructions, and optionally an internal bus 1003 and a network interface 1004 .
  • the memory 1002 may include a memory 10021, such as a high-speed random access memory (Random-Access Memory, RAM), and may also include a non-volatile memory 10022 (Non-Volatile Memory), such as at least one disk memory, etc.; processing
  • the device 1001, the network interface 1004 and the memory 1002 can be connected to each other through an internal bus 1003, and the internal bus 1003 can be an ISA (Industry Standard Architecture, industry standard architecture) bus, a PCI (Peripheral Component Interconnect, peripheral component interconnect standard) bus Or EISA (Extended Industry Standard Architecture, Extended Industry Standard Architecture) bus, etc.; the internal bus 1003 can be divided into address bus, data bus, control bus, etc., for the convenience of representation, only a bidirectional arrow is used in FIG.
  • ISA Industry Standard Architecture, industry standard architecture
  • PCI Peripheral Component Interconnect, peripheral component interconnect standard
  • EISA Extended Industry Standard Architecture
  • the processor 1001 executes the execution instructions stored in the memory 1002, the processor 1001 executes the method in any one of the embodiments of the present invention.
  • the processor reads the corresponding execution instructions from the non-volatile memory into the memory and then executes them, and also obtains the corresponding execution instructions from other devices, so as to form a logic level Equipment failure prediction device.
  • the processor executes the execution instructions stored in the memory, so as to implement the equipment failure prediction method provided in any embodiment of the present invention through the executed execution instructions.
  • a processor may be an integrated circuit chip with signal processing capabilities.
  • each step of the above-mentioned method can be completed by a hardware integrated logic circuit in a processor or an instruction in the form of software.
  • the above-mentioned processor can be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; it can also be a digital signal processor (Digital Signal Processor, DSP), dedicated integrated Circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
  • DSP Digital Signal Processor
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • a general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
  • Embodiments of the present invention further provide a computer-readable storage medium, including execution instructions.
  • execution instructions When a processor of an electronic device executes the execution instructions, the processor executes the method provided in any one of the embodiments of the present invention.
  • the electronic device may be the electronic device shown in FIG. 10 ; the execution instruction is a computer program corresponding to a method for predicting equipment failure.
  • embodiments of the present invention may be provided as a method or a computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or a combination of software and hardware.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Probability & Statistics with Applications (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Mathematics (AREA)
  • Algebra (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Medical Informatics (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了一种预测方法、装置、可读介质及电子设备，方法包括：确定目标设备的特征信息和目标设备对应的检测点数据信息；基于目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和非共享数据的概率分布模型；根据特征数据概率分布模型和检测点数据的概率分布模型，确定非共享数据的权重；根据非共享数据、非共享数据的权重和非共享数据对应的设备故障标签，建立联邦学习模型；根据联邦学习模型进行目标设备的设备故障预测。本发明提供的技术方案，可将非共享数据迁移到目标设备上，建立目标设备的特征数据和设备故障之间的关系，无需共享设备之间的特征数据，确保了数据安全。

Description

预测方法、装置、可读介质及电子设备 技术领域
本发明涉及能源和物联网技术领域，尤其涉及一种设备故障预测方法和装置、设备状态预测方法和装置、可读介质及电子设备。
背景技术
智能制造系统中的工业设备一旦发生故障，将会影响智能制造系统的正常工作，造成十分严重的影响，因此通常需要对工业设备进行设备故障预测。目前，在对某一工业设备进行预测时，由于该工业设备的运行数据没有标签数据，因此需要采集若干个其他工业设备的历史运行数据以及历史运行数据对应的标签数据，采用机器学习方法建立这些历史运行数据与设备故障预测之间的映射关系，得到设备故障预测模型，使用设备故障预测模型实现该工业设备的设备故障预测。但是，上述技术方案为了学习到有效的模型，在对多个工业设备的设备故障预测时，可能需要共享工业设备的历史运行数据以及历史运行数据对应的标签数据，因此导致工业设备之间的数据安全性较低。
设备预测性维护是根据设备运行状态的特征信息预测设备失效的概率或设备的剩余使用寿命等。其中基于数据驱动的方法是使用设备历史运行数据，包括设备各个传感器的量测，通过机器学习方法建立这些量测与设备发生故障概率之间的映射关系。为了学习到有效的模型，通常需要高质量的标注数据，也就是需要大量的故障标签数据；单台设备的故障数据有限，需要联合多台同类型的设备获得足够的数据。已有的技术方案能够解决设备自身故障预测的问题，但是对不同设备之间数据共享，目前已经有的技术方案，是通过直接共享物联网中各个设备的数据，并未考虑物联网中各设备之间保护数据隐私的需求。
此外，进行设备状态的自动监测，已经成为保障设备正常运行的一种重要技术手段。设备状态预测是根据设备运行数据预测设备故障的概率或设备的剩余使用寿命。目前，对于难以获得设备状态标签的待预测设备，通过获取其他设备的历史运行数据以及历史运行数据对应的设备状态标签数据，从而获取大量的标注数据，之后基于机器学习方法建立这些历史运行数据与设备状态标签数据之间的映射关系，获得设备状态预测模型，基于设备状态预测模型，从而实现对待预测设备的设备状态预测。但是，不同设备所产生的历史运行数据的数据分布有所差异，导致利用上述技术方案所得到的设备状态预测模型的精度相对较低。
发明内容
本发明针对现有技术中存在的上述技术问题提供了一种设备故障预测方法、装置、计算机可读存储介质及电子设备，可将非共享数据迁移到目标设备上，建立目标设备的特征数据和设备故障之间的关系，无需共享的设备之间的特征数据，确保了数据安全。
第一方面，本发明提供了一种设备故障预测方法，包括：确定目标设备的特征信息和目标设备对应的检测点数据信息；基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和具有非共享数据的检测点数据的概率分布模型；根据所述特征数据概率分布模型和所述检测点数据的概率分布模型，确定所述非共享数据的权重；根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，建立联邦学习模型；根据所述联邦学习模型进行所述目标设备的设备故障预测。
第二方面，本发明提供了一种设备故障预测装置，包括：信息确定模块，用于确定目标设备的特征信息和目标设备对应的检测点数据信息；概率模型确定模块，用于基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和具有非共享数据的检测点数据的概率分布模型；权重确定模块，用于根据所述特征数据概率分布模型和所述检测点数据的概率分布模型，确定所述非共享数据的权重；模型建立模块，用于根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签， 建立联邦学习模型；预测模块，用于根据所述联邦学习模型进行所述目标设备的设备故障预测。
第三方面，本发明实施例提供了一种设备故障预测方法，该设备故障预测方法包括：根据目标设备属性，获取用于对目标设备建立预测模型的训练数据集，其中所述数据集中的样本数据为共享数据；计算所述训练数据集中每条样本数据的权重；利用所述权重训练得到目标设备的故障预测局部模型；基于所述故障预测局部模型与联合学习算法，建立联合模型；根据所述联合模型对所述目标设备进行故障预测。
第四方面，本发明实施例提供了一种设备故障预测装置，该设备故障预测装置包括：数据获取模块、权重计算模块、局部模型训练模块、联合模型建立模块和故障预测模块，其中，所述数据获取模块，用于根据目标设备属性，获取用于对目标设备建立预测模型的训练数据集，其中所述数据集中的样本数据为共享数据；所述权重计算模块，用于计算所述训练数据集中每条样本数据的权重；所述局部模型训练模块，用于利用所述权重训练得到目标设备的故障预测局部模型；所述联合模型建立模块，用于基于所述故障预测局部模型与联合学习算法，建立联合模型；所述故障预测模块，用于根据所述联合模型对所述目标设备进行故障预测。
第五方面，本发明提供了一种设备状态预测方法，包括：获取目标设备对应的至少两个特征数据、至少两个参考设备分别对应的至少两个模型训练数据以及各个所述模型训练数据分别对应的设备状态标签，所述目标设备和所述参考设备的设备种类相同；根据各个所述特征数据，确定所述模型训练数据对应的误差权重；根据各个所述模型训练数据、各个所述模型训练数据分别对应的设备状态标签以及各个所述模型训练数据分别对应的误差权重进行模型训练，以确定设备状态预测模型，所述设备状态预测模型用于所述目标设备的设备状态预测。
第六方面，本发明提供了一种设备状态预测装置，包括：获取模块，用于获取目标设备对应的至少两个特征数据、至少两个参考设备分别对应的至少两个模型训练数据以及各个所述模型训练数据分别对应的设备状态标签，所述目标设备和所述参考设备的设备种类相同；权重确定模块，用于根据各个所述特征数据，确定所述模型训练数据对应的误差权重；训练模块，用于根据各个所述模型训练数据、各个所述模型训练数据分别对应的设备状态标签以及各个所述模型训练数据分别对应的误差权重进行模型训练，以确定设备状态预测模型，所述设备状态预测模型用于所述目标设备的设备状态预测。
第七方面，本发明提供了一种计算机可读存储介质，包括执行指令，当电子设备的处理器执行所述执行指令时，所述处理器执行如第一方面中任一所述的方法或如第二方面中任一所述的方法。
第八方面，本发明提供了一种电子设备，包括处理器以及存储有执行指令的存储器，当所述处理器执行所述存储器存储的所述执行指令时，所述处理器执行如第一方面中任一所述的方法或如第二方面中任一所述的方法。
本发明提供了一种设备故障预测方法和装置、设备状态预测方法和装置、计算机可读存储介质及电子设备，该方法通过确定目标设备的特征信息和目标设备对应的检测点数据信息，然后，基于目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和具有非共享数据的检测点数据的概率分布模型，然后根据特征数据概率分布模型和检测点数据的概率分布模型，确定非共享数据的权重，然后，根据非共享数据、非共享数据的权重和非共享数据对应的设备故障标签，建立联邦学习模型，最后，根据联邦学习模型进行目标设备的设备故障预测。本发明提供的技术方案，通过将非共享数据迁移到目标设备上，从而建立目标设备的特征数据和设备故障之间的关系，无需共享设备之间的特征数据，确保了数据安全。
上述的非惯用的优选方式所具有的进一步效果将在下文中结合具体实施方式加以说明。
附图说明
为了更清楚地说明本发明实施例或现有的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明中记载的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为本发明一实施例提供的一种设备故障预测方法的流程示意图；
图2为本发明一实施例提供的另一种设备故障预测方法的流程示意图；
图3为本发明一实施例提供的一种设备故障预测方法的流程示意图；
图4为本发明一实施例提供的另一种设备故障预测方法流程示意图；
图5为本发明一实施例提供的一种设备状态预测方法的流程示意图；
图6为本发明一实施例提供的另一种设备状态预测方法的流程示意图；
图7为本发明一实施例提供的一种设备故障预测装置的结构示意图；
图8为本发明一实施例提供的一种设备故障预测装置的结构示意图；
图9为本发明一实施例提供的一种设备状态预测装置的结构示意图；
图10为本发明一实施例提供的一种电子设备的结构示意图。
具体实施方式
为使本发明的目的、技术方案和优点更加清楚，下面将结合具体实施例及相应的附图对本发明的技术方案进行清楚、完整地描述。显然，所描述的实施例仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
图1为本发明一实施例提供的一种设备故障预测方法的流程示意图。如图1所示，本发明实施例提供了一种设备故障预测方法，包括如下多个步骤：
步骤101，确定目标设备的特征信息和目标设备对应的检测点数据信息。
具体地，目标设备为智能制造系统的任意一种工业设备，本发明实施例对此不作具体限定，优选地，目标设备为能源设备，比如，燃气锅炉、内燃机、汽轮机、热电联产设备、光伏设备等。
具体地，目标设备的特征信息指的是目标设备的特征数据的属性或功能。
具体地，检测点数据信息指的是检测点数据的属性或功能。
需要说明的是，检测点数据包括非共享数据。
步骤102，基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和具有非共享数据的检测点数据的概率分布模型。
具体地，目标设备的特征数据有多个，每个特征数据包括若干个特征分别对应在目标设备的特征值，其中，若干个特征为设备故障的影响因素，具体需要结合实际场景确定，比如，目标设备为燃气锅炉，则若干个特征包括但不限于燃气流量、燃气温度、排烟温度、烟气流量、燃气压力、开停机状态、烟气湿度以及烟气压力等。
具体地，特征数据服从连续概率分布，因此，特征数据概率分布模型可以是正态分布模型，也可以是指数分布模型，本发明实施例对此不做限定，优选地，特征数据概率分布模型是由多个正态分布模型混合成的混合高斯模型，其中，正态分布模型又称高斯分布模型。
具体地，检测点数据的概率分布模型与特征数据概率分布模型相同，优选地，可以是混合高斯模型，当然，也可以是其他概率分布模型，具体需要结合实际情况确定，此处不做具体限定。
本发明一实施例中，所述基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和具有非共享数据的检测点数据的概率分布模型确定所述特征数据概率分布模型，包括：基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据和具有非共享数据的检测点数据；根据特征数据参 数模型计算所述特征数据的数据分布，并将确定好参数的特征数据参数模型确定为所述特征数据概率分布模型；根据检测点数据参数模型计算所述检测点数据的数据分布，并将确定好参数的检测点数据参数模型确定为所述检测点数据的概率分布模型。
具体地，特征数据参数模型可以是高斯模型也可以是混合高斯模型，优选混合高斯模型。检测点数据参数模型与特征数据参数模型相同，本发明实施例对此不做过多赘述。
具体地，特征数据概率分布模型是通过将各个特征数据代入到混合高斯模型中进行计算来确定模型参数，其中计算方法可以是EM算法，将确定好模型参数的混合高斯模型确定为特征数据概率分布模型。
检测点数据的概率分布模型，是通过将多个检测点数据代入到混合高斯模型中进行计算来确定模型参数，其中计算方法可以是EM算法，将确定好模型参数的混合高斯模型确定为检测点数据的概率分布模型。
步骤103，根据所述特征数据概率分布模型和所述检测点数据的概率分布模型，确定所述非共享数据的权重。
该实施例中，通过非共享数据的权重，从而将检测点的非共享数据迁移到目标设备换上，换言之，建立了检测点和目标设备之间的数据关联，检测点不会直接获取到目标设备的特征数据，即目标设备和检测点之间不存在数据共享，从而确保了数据安全。
本发明一个实施例中，所述根据所述特征数据概率分布模型和所述检测点数据的概率分布模型，确定所述非共享数据的权重，包括：根据所述特征数据概率分布模型，确定所述非共享数据的目标设备分布概率；根据所述检测点数据的概率分布模型，确定所述非共享数据的检测点分布概率；将所述非共享数据的目标设备分布概率和所述非共享数据的检测点分布概率的比值，确定为所述非共享数据的权重。
该实施例中，通过将非共享数据的目标设备分布概率和非共享数据的检测点分布概率的比值，确定非共享数据的权重，从而将检测点上的非共享数据迁移到目标设备上，同时无需共享目标设备的特征数据，确保了数据安全。
具体地，将非共享数据代入特征数据概率分布模型中，特征数据概率分布模型输出的值即为非共享数据的目标设备分布概率。
具体地，将非共享数据代入检测点数据的概率分布模型，检测点数据的概率分布模型输出的值即为共享数据的检测点分布概率。
步骤104，根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，建立联邦学习模型。
该实施例中，通过非共享数据、非共享数据的权重和非共享数据对应的设备故障标签，建立联邦学习模型，实现将非共享数据到目标设备的迁移，无需共享目标设备的特征数据，从而确保了数据安全。
需要说明的是，目标设备的特征数据和检测点数据分布在物联网中的不同的检测点，共享数据进行模型训练会产生数据安全问题，本发明实施例通过检测点数据中的非共享数据、非共享数据的权重以及非共享数据对应的设备故障标签，实现将非共享数据迁移到目标设备上，使得检测点之间不存在数据共享，避免了直接共享数据带来的数据安全问题。其中，检测点是能进行数据处理以及数据交互的节点，包括但不限于边缘服务器、边缘网关以及边缘控制器中的任意一种或多种。
具体地，设备故障标签可以是设备运行状态的特征信息预测设备失效的概率，比如，设备故障标签可以是故障类型，还可以是故障程度，本发明实施例对此不作限定，具体需要结合实际需求确定。需要说明的是，本发明实施例并不意图对每个非共享数据分别对应的设备故障标签的获取方法进行限定，也可以是人工标注，也可以是规则标注，还可以是聚类标注，上述标注方法均为现有技术，此处不做过多赘述。值得注意的是，本发明实施例提供的设备故障预测方法，也可以用来预测设备剩余使用寿命，具体地，通过设备剩余使用寿命替换设备故障标签即可。
在一个实施例中，所述根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，建立联邦学习模型，包括：根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，确定检测点局部模型；根据至少两个所述检测点局部模型，建立联邦学习模型。
具体地，基于非共享数据、非共享数据对应的权重和设备故障标签进行模型训练，确定检测点局部模型，然后，对多个检测点局部模型进行融合，建立联邦学习模型。
具体地，非共享数据以及非共享数据的权重能够实现将检测点的非共享数据迁移到目标设备上，后续，通过多个非共享数据各自的权重调整模型参数，从而使得调整后的模型参数能够反映出目标设备的特征数据和设备故障之间的关系，不涉及到检测点数据和目标设备的特征数据的共享，从而确保了数据安全。
本发明的一个实施例中，所述根据至少两个所述检测点局部模型，建立联邦学习模型，包括：基于联合学习算法，对至少两个所述检测点局部模型进行反复迭代，并基于至少两个所述检测点的迭代好的检测点局部模型，建立联邦学习模型。
具体地，联合学习算法指的是将各个检测点局部模型中的检测点局部模型参数发送到目标设备的检测点，目标设备的检测点对各个检测点局部模型参数进行平均或加权平均得到联邦学习模型参数，基于联邦学习参数进行检测点局部模型更新迭代。
具体地，通过将各个检测点的检测点局部模型的模型参数发送到目标设备的检测点，目标设备的检测点将来自各检测点的检测点局部模型的模型参数进行融合得到联邦学习模型参数，并将融合后的联邦学习模型参数再分发给各个检测点的检测点局部模型，之后根据各个检测点的非共享数据和融合后的联合模型参数进行本地训练得到更新后的模型参数，将更新后的模型参数发送给目标设备的检测点，按照上述方法反复迭代，直到迭代次数满足预设次数，或者，迭代后的检测点局部模型的模型误差满足预设值。将迭代好的检测点局部模型参数进行融合以得到联邦学习模型参数，进而得到联邦学习模型。其中，各个检测点局部模型的模型参数的融合的方式可以是平均也可以是加权平均，本发明实施例对此不做限定。
具体地，检测点局部模型可以是神经网络模型，也可以是回归模型，具体需要结合实际需求确定。
具体地，可通过如下方式对检测点局部模型进行迭代：
A1、根据多个非共享数据、每个非共享数据分别对应的设备故障标签以及每个非共享数据分别对应的权重进行模型训练，以确定检测点局部模型；
A2、判断检测点局部模型的模型误差是否满足迭代条件，若是，则将检测点局部模型确定为最终模型发送给目标设备的检测点，若否，则执行A3；
A3、将检测点局部模型的模型参数发送给目标设备的检测点；
A4、接收目标设备的检测点发送的融合后的模型参数，并根据多个非共享数据、多个非共享数据分别对应的设备故障标签以及多个非共享分别对应的权重，对融合后的模型参数进行调整，以确定调整后的模型参数，并将调整后的模型参数替换检测点局部模型的模型参数，执行A2。
步骤S105，根据至少两个所述检测点的局部模型，建立联邦学习模型。
具体地，采集目标设备的当前特征数据，将当前特征数据代入联邦学习模型中，即可判断目标设备是否故障。
为了更好的理解目标设备的检测点和非共享数据的检测点之间的数据处理过程，举例来说，假设3个非共享数据的检测点设为A、B以及C，目标设备的节点设为D，以A为例进行说明，A根据混合高斯模型确定好非共享数据的概率分布模型，并接收D发送来的特征数据概率分布模型，从而确定非共享数据对应的权重，B以及C按照上述A相似的处理过程，得到B及C中的每个非共享数据分别对应的权重，之后，A、B、C分别根据各个非共享数据、非共享数据对应的设备故障标签以及每个非共享数据分别对应的权重进行模型训练，以得到检测点局部模型，之后，D对A、B、C发来的检测点局部模型进行联邦学习，即可得 到目标设备的联邦学习模型。
通过以上技术方案可知，本发明实施例至少存在以下有效效果：通过特征数据概率分布模型和非共享数据的概率分布模型确定非共享数据的权重，以得到目标设备的检测点局部模型，建立非共享数据的检测点和目标设备之间的数据关联，将非共享数据迁移到目标设备上，同时，基于非共享数据建立联邦学习模型，利用该联邦学习模型实现目标设备的设备故障预测，非目标设备和目标设备之间无需共享设备的特征数据，从而确保数据安全。
图1所示仅为本发明所述方法的基础实施例，在其基础上进行一定的优化和拓展，还能够得到所述方法的其他优选实施例。
为了更加清楚的说明本发明的技术方案，请参考图2。图2为本发明一实施例提供的另一种设备故障预测方法的流程示意图。本实施例在前述实施例的基础上，结合具体应用场景进行进一步的叙述。如图2所示，该设备故障预测方法具体可以包括如下各个步骤：
步骤201，确定目标设备的特征信息和目标设备对应的检测点数据信息。
具体地，假设目标设备对应在物联网中的检测点为n，有N个非共享数据的检测点，分别对应在物联网中的检测点分别为n1、n2、……、nN，n1、n2、……、nN分别能够获取到非共享数据，n能够获取到目标设备所有的特征数据。需要说明的是，n1、n2、……、nN得到检测点局部模型的过程相似，下述仅以n1的处理过程进行描述。
步骤202，基于所述目标设备的特征信息和混合高斯模型，计算所述特征信息对应的特征数据的数据分布，并将确定好模型参数的混合高斯模型确定为所述特征数据概率分布模型。
n根据高斯混合模型，将各个特征数据代入混合高斯模型中对各个特征数据计算，从而得到各个特征数据概率分布模型，并将各个特征数据概率分布模型分别送给n1、n2、……、nN。
n1、n2、……、nN得到非共享数据的概率分布模型的过程相似，下述以n1的处理过程进行描述。
步骤203，基于目标设备对应的检测点数据信息和混合高斯模型计算所述检测点数据信息对应的具有非共享数据的检测点数据的数据分布，并将确定好模型参数的混合高斯模型确定为所述非共享数据的概率分布模型。
n1获取多个非共享数据，将非共享数据代入混合高斯模型进行计算，得到n1的非共享数据的概率分布模型。
步骤204，根据所述特征数据概率分布模型，确定所述非共享数据的目标设备分布概率，根据所述检测点数据的概率分布模型，确定所述非共享数据的检测点分布概率。
n1接收到特征数据对应的概率分布模型，将其获得的非共享数据代入特征数据概率分布模型中进行计算，得到非共享数据的目标设备分布概率。
n1将非共享数据代入到非共享数据的概率分布模型中进行计算，得到非共享数据的检测点分布概率。
步骤205，将所述非共享数据的目标设备分布概率和所述非共享数据的检测点分布概率的比值，确定为所述非共享数据的权重。
n1通过将非共享数据相对于目标设备的分布概率和相对于非共享数据的检测点的分布概率的概率比值确定为非共享数据的权重，基于相同的方法，得到每个非共享数据的权重，这样就可以实现建立起目标设备和非共享数据的检测点之间的数据联系。
步骤206，根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，确定所述检测点局部模型。
n1根据非共享数据、非共享数据的权重及非共享数据对应的设备故障标签进行模型训练，确定目标设备的检测点局部模型。n2、……、nN按照n1相似的方法分别确定出目标设备的检测点局部模型。
步骤207，基于联合学习算法，对至少两个所述检测点的检测点局部模型进行反复迭代，并基于至少两个所述检测点的迭代好的检测点局部模型，建立联邦学习模型。
n与n1、n2、……、nN，采用联合学习算法，对n1、n2、……、nN分别训练得到的检测点局部模型进行模型迭代后发给n，n对n1、n2、……、nN分别训练得到的迭代好的检测点局部模型进行模型平均，建立目标设备的联邦学习模型。
步骤208，将所述联邦学习模型进行所述目标设备的设备故障预测。
n根据联邦学习模型进行目标设备的设备故障预测。
通过以上技术方案可知，本实施例存在的有益效果是：通过非共享数据的概率分布模型与特征数据概率分布模型的比值，得到非共享数据的权重，建立非共享数据的检测点和目标设备之间的数据关联，将非共享数据迁移到目标设备上，通过目标设备和非共享数据在物联网中的检测点实现目标设备的联邦学习模型的建立，从而实现目标设备和非共享数据的检测点之间不涉及到特征数据的共享，确保了数据安全。
图3为本发明一实施例提供的一种设备故障预测方法的流程示意图。如图3所示，该设备故障预测方法包括：
步骤301，根据目标设备属性，获取用于对目标设备建立预测模型的训练数据集，其中所述数据集中的样本数据为共享数据；
步骤302，计算所述训练数据集中每条样本数据的权重；
步骤303，利用所述权重训练得到目标设备的故障预测局部模型；
步骤304，基于所述故障预测局部模型与联合学习算法，建立联合模型；
步骤305，根据所述联合模型对所述目标设备进行故障预测。
在该实施例中，训练数据集中的样本数据包含目标设备的特征数据和样本设备的特征数据、样本设备的故障数据；样本设备为与目标设备相关或相似的设备。故障预测局部模型和联合模型均为设备的特征数据与设备发生故障的关系。不同设备数据分布在物联网不同的节点，共享数据训练模型会产生数据安全问题，该实施例采用一种基于联合学习方式的样本迁移方法，用于设备预测性维护，可以将在多台设备上采集到的数据联合学习，并迁移到目标设备上，用来训练针对目标设备的预测性维护模型，实现多方联合学习，保证各方数据不出本地，避免了直接共享数据带来的数据安全问题。
在本发明一个实施例中，所述计算所述训练数据集中每条样本数据的权重，包括针对每台所述样本设备：将所述样本设备的特征数据与所述目标设备的特征数据进行区分；根据区分后的特征数据，训练分类模型；根据训练后的分类模型，计算所述样本设备每条特征数据的权重。
由于样本设备和待预测设备可以分别分布物联网中的任何节点，为保证数据隐私，样本设备和待预测设备的数据不能共享。因此可以采用基于联合学习的分类模型，如在该实施例中使用基于联合学习的XGBoost模型。
在本发明一个实施例中，所述计算所述训练数据集中每条样本数据的权重，包括针对每台所述样本设备：将所述样本设备的特征数据标记为第一数据，将所述待预测设备的特征数据标记为第二数据；根据所述第一数据和所述第二数据，训练分类模型，所述分类模型为基于联合学习的分类模型；根据训练后的分类模型，计算所述样本设备每条特征数据的权重，所述权重的计算公式为：
Figure PCTCN2021101314-appb-000001
其中，ω i为所述第一数据中第i条数据的权重，P 1i为所述第i条数据属于所述样本设备的概率，P 2i为所述第i条数据属于所述目标设备的概率。
在本发明一个实施例中，所述利用所述权重训练得到目标设备的故障预测局部模型，包括：根据每台所述样本设备的特征数据、所述样本设备每条特征数据的权重和每台所述样本设备的故障数据，使用神经网络分别在所述样本设备有权重的所述训练数据集上训练得到所述目标设备的故障预测局部模型。
在本发明一个实施例中，所述基于所述故障预测局部模型与联合学习算法，建立联合模型，包括：根据所述故障预测局部模型，使用联合学习算法反复迭代，获得所述样本设备在所述训练数据集上关于所述目标设备的联合模型。
在本发明一个实施例中，所述目标设备和各个所述样本设备均为物联网中的边缘节点，所述目标设备的特征数据不对其他所述样本设备暴露，每个所述样本设备的特征数据和故障数据不对其他所述样本设备和所述目标设备暴露。
图4为本发明一实施例提供的另一种设备故障预测方法流程示意图。如图4所示，该设备故障预测方法包括：
步骤401，获取每台样本设备的传感器特征数据，以及相应的故障数据；采集目标设备的传感器特征数据；
步骤402，分别为目标设备和样本设备的特征数据打标签，得到标签数据；
步骤403，基于步骤402的标签数据，采用联合学习的方式训练分类器(二分类)；
步骤404，根据步骤403的分类器，计算样本设备特征数据的权重；
步骤405，重复以上步骤计算各样本设备特征数据的权重；
步骤406，基于各个样本设备的数据，采用联合学习的方式建立传感器特征数据和设备发生故障之间的关系；
步骤407，将步骤406的模型用于目标设备发生故障概率的预测。
假定具有设备A、B、C等各个测点的特征数据，以及相应的故障标注，也就是具有设备A、B、C的所有特征数据和故障数据(标签数据)；同时也具有设备D相同测点的特征数据，也就是具有设备D的所有特征数据，不具备故障数据，需要根据这些特征数据预测设备D的故障发生概率。
目标是使用设备A、B、C的特征数据、故障数据和设备D的特征数据学习到一个预测设备D发生故障的概率模型，用于设备D的故障预测。
需要注意的是在此过程中，假定A、B、C、D等均为物联网中的边缘节点，各方都有数据隐私和安全的需求，在训练过程中自身特征数据和故障数据均不能对外暴露。
在该实施例中，设备故障预测方法包括：
1.收集训练数据集训练数据集包括设备A、B、C的所有特征数据以及故障数据，还有设备D的特征数据；
2.计算设备A、B、C每条特征数据的权重，以下以设备A为例，B，C情况相同，计算步骤如下：
a)为设备A和设备D的特征数据打标签，假定设备A的标签为0，设备D的标签为1；
b)基于上述设备A和设备D的标签数据，训练分类器(二分类)(本例中使用基于联合学习的XGBoost模型，实际中不限使用XGBoost，可以使用任何基于联合学习的概率分类模型)；
c)对于设备A中的每条数据x，根据上述训练的二分类分类器，计算数据x属于设备A的概率PA(x)，以及属于设备D概率PD(x)；计算数据x权重，ω(x)＝PD(x)/PA(x)；
3.训练模型
a)由以上步骤得到设备A、B、C的特征数据和故障数据，以及每条特征数据对应的权重；
b)训练局部模型，使用神经网络或相关回归算法分别在上述设备A、B、C的有权重的数据集上训练得到设备D的故障预测局部模型(具体训练过程为标准过程，不再赘述)；
c)计算联合模型，根据上述在设备A、B、C数据集上训练得到的设备D的故障预测局部模型；使用模型平均(联合学习算法)反复迭代，计算设备A、B、C数据集上关于设备D故障预测的联合模型也就是故障预测模型；在该实施例中，反复迭代的次数由模型的预设精度和预设迭代次数决定。
4.将得到的联合模型用于设备D的故障预测。
本实施例中暂假定设备A、B、C三台设备参与联合训练，在实际中，联合设备的数量没有限制。
值得注意的是，本发明中提到的设备故障数据，同样可以替换为设备剩余使用寿命，从而对设备的剩余使用寿命进行预测。
图5为本发明一实施例提供的一种设备状态预测方法的流程示意图。如图5所述，该设备状态预测方法包括如下各个步骤：
步骤501，获取目标设备对应的至少两个特征数据、至少两个参考设备分别对应的至少两个模型训练数据以及各个所述模型训练数据分别对应的设备状态标签，所述目标设备和所述参考设备的设备种类相同。
具体地，可以根据接入网络的参考设备上安装的传感器，获取参考设备的运行数据，根据参考设备的运行数据，从而确定出参考设备对应的多个模型训练数据。需要说明的是，基于传感器采集的数据获取模型训练数据为现有技术，本实施例对此不作过多赘述。
具体地，可以根据接入网络的目标设备安装的传感器，获取目标设备的运行数据，根据目标设备的运行数据，从而确定出目标设备对应的多个特征数据。需要说明的是，基于传感器采集的数据获取特征数据为现有技术，本实施例对此不作过多赘述。
具体地，模型训练数据包括若干个特征以及每个特征分别对应在参考设备的一个或多个特征值，优选一个，其中，若干个特征为对设备状态具有影响的影响因素。模型训练数据和特征数据所对应的若干个特征相同，特征对应的特征值可能不同。这里，模型训练数据分别对应一个设备状态标签。需要说明的是，本发明实施例并不意图对参考设备的数量进行限定，具体需要结合实际场景确定。
在一个实施例中，设备状态标签包括但不限于设备故障信息或设备的剩余使用寿命。需要说明的是，本发明实施例并不意图对每个模型训练数据分别对应的设备状态标签的获取方法进行限定，也可以是人工标注，也可以是规则标注，还可以是聚类标注，上述标注方法均为现有技术，此处不做过多赘述。
需要说明的是，参考设备和目标设备的设备种类相同，从而确保能够对参考设备和目标设备进行数据关联的价值，进而确保后续确定出的设备状态预测模型的参考价值。比如，参考设备和目标设备均是燃气锅炉。
还需要说明的是，目标设备的特征数据没有设备状态标签，无法利用自身的数据完成模型训练，因此需要借助参考设备的模型训练数据，得到其对应的设备状态预测模型。
步骤502，根据各个所述特征数据，确定所述模型训练数据对应的误差权重。
在一个实施例中，具体可通过如下方法确定模型训练数据对应的误差权重：
从各个特征数据中确定出模型训练数据对应的替换数据；确定模型训练数据在其所属的参考设备对应的各个模型训练数据中的第一分布信息以及替换数据在各个特征数据中的第二分布信息；根据第一分布信息和第二分布信息，确定模型训练数据对应的权重误差。
为了能够确定出参考设备的模型训练数据对应的误差权重，了解参数设备和目标设备之间的数据关系，需要从目标设备的多个特征数据中找到模型训练数据对应的替换数据，这里，替换数据优选多个特征数据中与模型训练数据相同的数据，若多个特征数据中不存在和模型训练数据相同的数据，此时，替换数据可以为多个特征数据中和模型训练数据最相似的数据，作为一种可能的情况，当多个特征数据中存在2个或2个以上可替换模型训练数据的特征数据时，此时，可以随机选择一个可替换模型训练数据的特征数据作为替换数据。
然后，确定模型训练数据在其所属的参考设备对应的各个模型训练数据中的第一分布信息。这里，第一分布信息指示了模型训练数据在其所属的参考设备对应的所有模型训练数据的分布情况，第一分布信息可以是分布概率，也可以是出现频率，此处不做具体限定。
然后，确定替换数据在各个特征数据中的第二分布信息。这里，第二分布信息指示了替换数据在目标设备的所有特征数据的分布情况，第二分布信息可以是分布概率，也可以是出现频率，此处不做具体限定。需要说明的是，第一分布信息和第二分布信息所包括的内容应 当统一，比如，第一分布信息和第二分布信息可以都包括分布概率，也可以都包括出现频率，此处不做具体限定。
需要说明的是，本实施例并不意图对确定分布信息的方法进行限定，任何现有技术中用于确定数据分布的方法即可，比如，可以是参数估计方法，也可以是非参数估计方法，还可以是半参数估计方法。
在一个实施例中，第一分布信息为第一分布概率，第二分布信息为第二分布概率，具体可通过如下方式确定第一分布概率及第二分布概率：
基于模型训练数据在其所属的参考设备对应的第一非参数估计方法，确定模型训练数据在其所属的参考设备对应的各个模型训练数据中的第一分布概率；基于第二非参数估计方法，确定特征数据在各个特征数据中的第二分布概率。
非参数估计方法不受总体分布的限制，不假定总体分布的具体形式，尽量从数据或样本本身获得所需要的信息，通过估计而获得分布的结构，对于减少偏差、提高预测精度以及了解数据的动态结构具有极其重要的作用，尤其是数据分布很不规则时，非参数估计方法相对于参数估计来说，更为精准，因此本实施例中采用非参数估计方法确定分布概率。
考虑到参考设备和目标设备分别对应的数据分布可能存在差异，因此，为了确保得到的分布概率的准确性，通过第一非参数估计方法确定参考设备对应的各个模型训练数据的概率密度函数，基于概率密度分布函数从而确定出参考设备对应的每个模型训练数据的第一分布概率。通过第二非参数估计方法确定目标设备对应的各个特征数据的概率密度函数，基于概率密度分布函数从而确定出目标设备对应的每个特征数据的第二分布概率。
考虑到设备的数据量和数据分布会存在差异，若采用相同的非参数估计方法，则可能会降低估算出的数据分布的准确性，因此，需要结合每个参考设备以及目标设备的数据量以及数据分布，选择合适的非参数估计方法进行数据分布的估算。当每个参考设备分别对应的各个模型训练数据的数据量，和目标设备对应的各个特征数据的数据量的差异较小时，各个参考设备分别对应的第一非参数估计方法和目标设备对应的第二非参数估计方法相同，作为一种可能的情况，第一非参数估计方法和第二非参数估计方法均为核密度估计方法。
在一个实施例中，当第一分布信息为第一分布概率，第二分布信息为第二分布概率时，具体可通过如下方式确定模型训练数据对应的误差权重：
针对每个模型训练数据，确定模型训练数据对应的替换数据的第二分布概率和模型训练数据的第一分布概率的分布概率比值，将分布概率比值确定为确定模型训练数据对应的误差权重。
在一个实施例中，当第一分布信息为第一出现频率，第二分布信息为第二出现频率时，具体可通过如下方式确定模型训练数据对应的误差权重：
针对每个模型训练数据，确定模型训练数据对应的替换数据的第二出现频率和模型训练数据的第一出现频率的出现频率比值，将出现频率比值确定为确定模型训练数据对应的误差权重。
步骤503，根据各个所述模型训练数据、各个所述模型训练数据分别对应的设备状态标签以及各个所述模型训练数据分别对应的误差权重进行模型训练，以确定设备状态预测模型，所述设备状态预测模型用于所述目标设备的设备状态预测。
在一个实施例中，设备状态预测模型可以是神经网络模型，也可以是回归模型，此处不做限定，具体需要结合实际情况确定。
在一个实施例中，本实施例提供的设备状态预测方法应用于目标设备对应的边缘服务器或本地服务器，从而降低云端服务器的计算压力。
具体地，设备状态预测模型以多个模型训练数据以及每个模型训练数据分别对应的设备状态标签数据作为训练数据，通过每个模型训练数据分别对应的误差权重调整设备状态预测模型的模型参数，从而使得调整后的模型参数能够反映出目标设备的特征数据和设备状态之间的联系，确保设备状态预测模型的模型精度。
在一个实施例中，通过可通过如下实现方式确定设备状态预测模型：
A1、根据将模型训练数据代入待预测模型中的预测结果以及模型训练数据对应的设备状态标签，确定模型训练数据分别对应的第一误差，根据模型训练数据对应的第一误差以及误差权重，以确定模型训练数据对应的第二误差；
A2、判断是否满足迭代次数或者各个模型训练数据分别对应的第二误差是否满足预设条件，如果是，则将待预测模型确定为设备状态预测模型，如果否，则执行A3；
A3、根据各个模型训练数据分别对应的第二误差，对待预测模型中的模型参数进行调整，以确定调整后的模型参数，并将待预测模型中的模型参数替换为调整后的模型参数，执行A1。
这里，各个模型训练数据分别对应的第二误差是否满足预设条件可以是各个模型训练数据分别对应的第二误差的均值是否小于阈值，也可以是各个模型训练数据分别对应的第二误差的标准方差是否小于阈值，此处不做具体限定。
具体地，预测结果可以理解为设备状态信息，第一误差为预测结果和设备状态标签之间的接近程度，举例来说，预测结果可以是故障或者不故障，当预测结果是故障，设备状态标签也是故障时，第一误差可以是0，否则为1。
通过以上技术方案可知，本实施例存在的有益效果是：通过确定多个参考设备分别对应的多个模型训练数据分别相对于目标设备的误差权重，从而将多个参考设备分别对应的多个模型训练数据迁移到目标设备上，后续，利用多个参考设备分别对应的多个模型训练数据、每个模型训练数据对应的设备状态标签以及每个模型训练数据对应的权重误差进行模型训练，得到用于预测目标设备的设备状态的设备状态预测模型，得到的设备状态预测模型由于综合考虑了模型训练数据对于目标设备的误差权重，从而具有相对较高的精度，后续基于设备状态预测模型进行目标设备的设备状态预测时，能够得到较为准确的目标设备的设备状态。
图5所示仅为本发明所述方法的基础实施例，在其基础上进行一定的优化和拓展，还能够得到所述方法的其他优选实施例。
图6为本发明一实施例提供的另一种设备状态预测方法的流程示意图。本实施例在前述实施例的基础上，结合应用场景进行了更加具体的描述。如图6所示，该设备状态预测方法具体包括以下步骤：
步骤601，获取目标设备对应的至少两个特征数据、至少两个参考设备分别对应的至少两个模型训练数据以及各个所述模型训练数据分别对应的设备状态标签，所述目标设备和所述参考设备的设备种类相同。
假设3个参考设备分别为锅炉A、锅炉B以及锅炉C，目标设备为锅炉D，锅炉A、锅炉B以及锅炉C分别对应有N个模型训练数据，每个模型训练数据分别对应有一个故障标签，故障标签可以是故障或者正常，锅炉D对应有n个特征数据，特征数据和模型训练数据分别对应的若干个特征相同，但是每个特征分别对应的特征值可能不同。
步骤602，基于所述模型训练数据在其所属的参考设备对应的第一非参数估计方法，确定所述模型训练数据在其所属的参考设备对应的各个所述模型训练数据中的第一分布概率。
假设锅炉A、锅炉B以及锅炉C分别对应的第一非参数估计方法均为核密度估计方法，针对锅炉A，基于核密度估计方法，确定锅炉A对应有N个模型训练数据的概率密度分布函数，针对锅炉A中的每个模型训练数据，基于锅炉A对应有N个模型训练数据的概率密度分布函数，确定该模型训练数据在锅炉A中的所有模型训练数据中的第一分布概率P1。锅炉B及锅炉C分别对应的每个模型训练数据的第一分布概率P1的计算方法和锅炉A相似，这里不做过多赘述。
步骤603，针对每个所述模型训练数据，从各个所述特征数据中确定出所述模型训练数据对应的替换数据；基于第二非参数估计方法，确定所述替换数据在各个所述特征数据中的第二分布概率。
针对锅炉A、锅炉B以及锅炉C分别对应的每个模型训练数据，从各个特征数据中确 定该模型训练数据对应的替换数据；基于核密度估算方法，确定锅炉D对应的n个特征数据的概率密度分布函数；针对每个替换数据，基于锅炉D对应的n个特征数据的概率密度分布函数，确定该替换数据在锅炉D中的所有特征数据中的第二分布概率P2。
步骤604，针对每个所述模型训练数据，确定所述模型训练数据对应的替换数据的第二分布概率和所述模型训练数据的第一分布概率的分布概率比值；将所述分布概率比值确定为确定所述模型训练数据对应的误差权重。
针对每个模型训练数据，将模型训练数据对应的替换数据的第二分布概率P2，与模型训练数据的第一分布概率P1的比值P2/P1，将P2/P1确定为该模型训练数据对应的误差权重W。
步骤605，根据各个所述模型训练数据、各个所述模型训练数据分别对应的设备状态标签以及各个所述模型训练数据分别对应的误差权重进行模型训练，以确定设备状态预测模型，所述设备状态预测模型用于所述目标设备的设备状态预测。
具体地，以锅炉A、锅炉B以及锅炉C分别对应的N个模型训练数据以及每个模型训练数据分别对应的故障标签进行设备状态预测模型的训练，基于每个模型训练数据分别对应的误差权重W调整模型参数，最终得到设备状态预测模型。
基于与本发明方法实施例提供的一种设备故障预测方法相同的构思，请参考图7。图7为本发明一实施例提供的一种设备故障预测装置的结构示意图。如图7所示，该设备状态故障预测装置，包括：
信息确定模块701，用于确定目标设备的特征信息和目标设备对应的检测点数据信息；
概率模型确定模块702，用于基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和具有非共享数据的检测点数据的概率分布模型；
权重确定模块703，用于根据所述特征数据概率分布模型和所述检测点数据的概率分布模型，确定所述非共享数据的权重；
模型建立模块704，用于根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，建立联邦学习模型；
预测模块705，用于根据所述联邦学习模型进行所述目标设备的设备故障预测。
本发明的一个实施例中，所述概率模型确定模块702，包括：数据确定单元、第一概率分布模型确定单元、第二概率分布模型确定单元；其中，数据确定单元，用于基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据和具有非共享数据的检测点数据；第一概率分布模型确定单元，用于基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据和具有非共享数据的检测点数据；第二概率分布模型确定单元，用于根据检测点数据参数模型计算所述检测点数据的数据分布，并将确定好参数的检测点数据参数模型确定为所述检测点数据的概率分布模型。
该实施例中，所述特征数据参数模型和/或检测点数据参数模型包括混合高斯模型。
本发明的一个实施例中，所述权重确定模块703，包括：第一分布概率确定单元、第二分布概率单元以及权重确定单元；其中，所述第一分布概率确定单元，用于根据所述特征数据概率分布模型，确定所述非共享数据的目标设备分布概率；所述第二分布概率单元，用于根据所述检测点数据的概率分布模型，确定所述非共享数据的检测点分布概率；所述权重确定单元，用于将所述非共享数据的目标设备分布概率和所述非共享数据的检测点分布概率的比值，确定为所述非共享数据的权重。
本发明的一个实施例中，所述模型建立模块704，包括：局部模型确定单元以及模型建立单元；所述局部模型确定单元，用于根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，确定检测点局部模型；模型建立单元，用于根据至少两个所述检测点的检测点局部模型，建立联邦学习模型。
本发明的一个实施例中，所述模型建立单元，用于基于联合学习算法，对至少两个所述 检测点的检测点局部模型进行反复迭代，并基于至少两个所述检测点的迭代好的检测点局部模型，建立联邦学习模型。
本发明的一个实施例中，所述检测点局部模型包括神经网络模型或回归模型。
图8为本发明一实施例提供的一种设备故障预测装置的结构示意图。如图8所示，该设备故障预测装置包括：数据获取模块801、权重计算模块802、局部模型训练模块803、联合模型建立模块804和故障预测模块805，其中，
所述数据获取模块801，用于根据目标设备属性，获取用于对目标设备建立预测模型的训练数据集，其中所述数据集中的样本数据为共享数据；
所述权重计算模块802，用于计算所述训练数据集中每条样本数据的权重；
所述局部模型训练模块803，用于利用所述权重训练得到目标设备的故障预测局部模型；
所述联合模型建立模块804，用于基于所述故障预测局部模型与联合学习算法，建立联合模型；
所述故障预测模块805，用于根据所述联合模型对所述目标设备进行故障预测。
在本发明一个实施例中，所述训练数据集中的所述样本数据包含目标设备的特征数据和样本设备的特征数据、样本设备的故障数据；所述样本设备为与所述目标设备相关或相似的设备。
在本发明一个实施例中，所述权重计算模块包括数据标记单元、数据分类单元和数据计算单元，其中，针对每台所述样本设备，所述数据标记单元，用于将所述数据获取模块获取的所述样本设备的特征数据与所述目标设备的特征数据进行区分；所述数据分类单元，用于根据所述数据标记单元区分后的特征数据，训练分类模型；所述数据计算单元，根据所述数据分类单元训练后的分类模型，计算所述样本设备每条特征数据的权重。
在本发明一个实施例中，针对每台所述样本设备：所述数据标记单元，具体用于将所述数据获取模块获取的所述样本设备的特征数据标记为第一数据，将所述数据获取模块获取的所述目标设备的特征数据标记为第二数据；所述数据分类单元，具体用于根据所述数据标记单元标记的所述第一数据和所述第二数据，训练分类模型，所述分类模型为基于联合学习的分类模型；所述数据计算单元，具体用于根据所述数据分类单元训练后的分类模型，计算所述样本设备每条特征数据的权重，所述权重的计算公式为：
Figure PCTCN2021101314-appb-000002
其中，ω i为所述第一数据中第i条数据的权重，P 1i为所述第i条数据属于所述样本设备的概率，P 2i为所述第i条数据属于所述目标设备的概率。
在本发明一个实施例中，所述局部模型训练模块具体用于根据每台所述样本设备的特征数据、所述样本设备每条特征数据的权重和每台所述样本设备的故障数据，使用神经网络分别在所述样本设备有权重的所述训练数据集上训练得到所述目标设备的故障预测局部模型。
在本发明一个实施例中，所述联合模型建立模块具体用于根据所述故障预测局部模型，使用联合学习算法反复迭代，获得所述样本设备在所述训练数据集上关于所述目标设备的联合模型。
在本发明一个实施例中，所述目标设备和各个所述样本设备均为物联网中的边缘节点，所述目标设备的特征数据不对其他所述样本设备暴露，每个所述样本设备的特征数据和故障数据不对其他所述样本设备和所述目标设备暴露。
基于与本发明方法实施例相同的构思，请参考图9。图9为本发明一实施例提供的一种设备状态预测装置的结构示意图。如图9所示，该设备状态预测装置，包括：
获取模块901，用于获取目标设备对应的至少两个特征数据、至少两个参考设备分别对应的至少两个模型训练数据以及各个所述模型训练数据分别对应的设备状态标签，所述目标设备和所述参考设备的设备种类相同；
权重确定模块902，用于根据各个所述特征数据，确定所述模型训练数据对应的误差权重；
训练模块903，用于根据各个所述模型训练数据、各个所述模型训练数据分别对应的设备状态标签以及各个所述模型训练数据分别对应的误差权重进行模型训练，以确定设备状态预测模型，所述设备状态预测模型用于所述目标设备的设备状态预测。
在一个实施例中，所述权重确定模块902，包括：替换数据确定单元、分布信息确定单元及重要程度确定单元；其中，所述替换数据确定单元，用于从各个所述特征数据中确定出所述模型训练数据对应的替换数据；所述分布信息确定单元，用于确定所述模型训练数据在其所属的参考设备对应的各个所述模型训练数据中的第一分布信息以及所述替换数据在各个所述特征数据中的第二分布信息；所述重要程度确定单元，用于根据所述模型训练数据的第一分布信息以及所述模型训练数据对应的替换数据的第二分布信息，确定所述模型训练数据对应的误差权重。
在一个实施例中，所述分布信息确定单元，包括：第一信息确定子单元以及第二信息确定子单元；其中，所述第一信息确定子单元，用于基于所述模型训练数据在其所属的参考设备对应的第一非参数估计方法，确定所述模型训练数据在其所属的参考设备对应的各个所述模型训练数据中的第一分布概率，将所述第一分布概率确定为第一分布信息；所述第二信息确定子单元，用于基于第二非参数估计方法，确定所述替换数据在各个所述特征数据中的第二分布概率，将所述第二分布概率确定为第二分布信息。
在一个实施例中，所述第一非参数估计方法和所述第二非参数估计方法均为核密度估计算法。
在一个实施例中，所述权重确定模块302，用于针对每个所述模型训练数据，确定所述模型训练数据对应的替换数据的第二分布概率和所述模型训练数据的第一分布概率的分布概率比值，将所述分布概率比值确定为确定所述模型训练数据对应的误差权重。
在一个实施例中，所述方法应用到所述目标设备对应的边缘服务器或本地服务器。
在一个实施例中，所述训练模块903，包括：训练单元、判断单元以及调整单元；其中，所述训练单元，用于根据将所述模型训练数据代入待预测模型中的预测结果以及所述模型训练数据对应的设备状态标签，确定所述模型训练数据对应的第一误差，根据所述模型训练数据对应的第一误差以及误差权重，以确定所述模型训练数据对应的第二误差；所述判断单元，用于判断是否满足迭代次数或者各个所述模型训练数据分别对应的第二误差是否满足预设条件，如果是，则将所述待预测模型确定为设备状态预测模型，如果否，则触发所述调整单元；所述调整单元，用于根据各个所述模型训练数据分别对应的第二误差，对所述待预测模型中的模型参数进行调整，以确定调整后的模型参数，并将所述待预测模型中的模型参数替换为所述调整后的模型参数，触发所述训练单元。
上述装置内的各模块和单元之间的信息交互、执行过程等内容，由于与本发明方法实施例基于同一构思，具体内容可参见本发明方法实施例中的叙述，此处不再赘述。
图10是本发明实施例提供的一种电子设备的结构示意图。在硬件层面，该电子设备包括处理器1001以及存储有执行指令的存储器1002，可选地还包括内部总线1003及网络接口1004。其中，存储器1002可能包含内存10021，例如高速随机存取存储器(Random-Access Memory，RAM)，也可能还包括非易失性存储器10022(Non-Volatile Memory)，例如至少1个磁盘存储器等；处理器1001、网络接口1004和存储器1002可以通过内部总线1003相互连接，该内部总线1003可以是ISA(Industry Standard Architecture，工业标准体系结构)总线、PCI(Peripheral Component Interconnect，外设部件互连标准)总线或EISA(Extended Industry Standard Architecture，扩展工业标准结构)总线等；内部总线1003可以分为地址总线、数据总线、控制总线等，为便于表示，图10中仅用一个双向箭头表示，但并不表示仅有一根总线或一种类型的总线。当然，该电子设备还可能包括其他业务所需要的硬件。当处理器1001执行存储器1002存储的执行指令时，处理器1001执行本发明任意一个实施例中 的方法。
在一种可能实现的方式中，处理器从非易失性存储器中读取对应的执行指令到内存中然后运行，也可从其它设备上获取相应的执行指令，以在逻辑层面上形成一种设备故障预测装置。处理器执行存储器所存放的执行指令，以通过执行的执行指令实现本发明任一实施例中提供的一种设备故障预测方法。
处理器可能是一种集成电路芯片，具有信号的处理能力。在实现过程中，上述方法的各步骤可以通过处理器中的硬件的集成逻辑电路或者软件形式的指令完成。上述的处理器可以是通用处理器，包括中央处理器(Central Processing Unit，CPU)、网络处理器(Network Processor，NP)等；还可以是数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。可以实现或者执行本发明实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
本发明实施例还提供了一种计算机可读存储介质，包括执行指令，当电子设备的处理器执行执行指令时，所述处理器执行本发明任意一个实施例中提供的方法。该电子设备具体可以是如图10所示的电子设备；执行指令是一种设备故障预测方法所对应计算机程序。
本领域内的技术人员应明白，本发明的实施例可提供为方法或计算机程序产品。因此，本发明可采用完全硬件实施例、完全软件实施例，或软件和硬件相结合的形式。
本发明中的多个实施例均采用递进的方式描述，多个实施例之间相同相似的部分互相参见即可，每个实施例重点说明的都是与其他实施例的不同之处。尤其，对于装置实施例而言，由于其基本相似于方法实施例，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。
还需要说明的是，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、商品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、商品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、商品或者设备中还存在另外的相同要素。
以上所述仅为本发明的实施例而已，并不用于限制本发明。对于本领域技术人员来说，本发明可以有各种更改和变化。凡在本发明的精神和原理之内所作的任何修改、等同替换、改进等，均应包含在本发明的权利要求范围之内。

Claims (26)

  1. 一种设备故障预测方法，其特征在于，包括：
    确定目标设备的特征信息和目标设备对应的检测点数据信息；
    基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和具有非共享数据的检测点数据的概率分布模型；
    根据所述特征数据概率分布模型和所述检测点数据的概率分布模型，确定所述非共享数据的权重；
    根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，建立联邦学习模型；
    根据所述联邦学习模型进行所述目标设备的设备故障预测。
  2. 根据权利要求1所述的方法，其特征在于，所述根据所述特征数据概率分布模型和所述检测点数据的概率分布模型，确定所述非共享数据的权重，包括：
    根据所述特征数据概率分布模型，确定所述非共享数据的目标设备分布概率；
    根据所述检测点数据的概率分布模型，确定所述非共享数据的检测点分布概率；
    将所述非共享数据的目标设备分布概率和所述非共享数据的检测点分布概率的比值，确定为所述非共享数据的权重。
  3. 根据权利要求1所述的方法，其特征在于，所述基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和具有非共享数据的检测点数据的概率分布模型，包括：
    基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据和具有非共享数据的检测点数据；
    根据特征数据参数模型计算所述特征数据的数据分布，并将确定好参数的特征数据参数模型确定为所述特征数据概率分布模型；
    根据检测点数据参数模型计算所述检测点数据的数据分布，并将确定好参数的检测点数据参数模型确定为所述检测点数据的概率分布模型。
  4. 根据权利要求3所述的方法，其特征在于，所述特征数据参数模型和/或检测点数据参数模型包括混合高斯模型。
  5. 根据权利要求1所述的方法，其特征在于，所述根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，建立联邦学习模型，包括：
    根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，确定检测点局部模型；
    根据至少两个所述检测点局部模型，建立联邦学习模型。
  6. 根据权利要求5所述的方法，其特征在于，所述根据至少两个所述检测点局部模型，建立联邦学习模型，包括：
    基于联合学习算法，对至少两个所述检测点局部模型进行反复迭代，并基于至少两个迭代好的检测点局部模型，建立联邦学习模型。
  7. 根据权利要求5所述的方法，其特征在于，所述检测点局部模型包括神经网络模型或回归模型。
  8. 一种设备故障预测方法，其特征在于，包括：
    根据目标设备属性，获取用于对目标设备建立预测模型的训练数据集，其中所述数据集中的样本数据为共享数据；
    计算所述训练数据集中每条样本数据的权重；
    利用所述权重训练得到目标设备的故障预测局部模型；
    基于所述故障预测局部模型与联合学习算法，建立联合模型；
    根据所述联合模型对所述目标设备进行故障预测。
  9. 根据权利要求8所述设备故障预测方法，其特征在于，所述训练数据集中的所述样本数据包含目标设备的特征数据和样本设备的特征数据、样本设备的故障数据；所述样本设备为与所述目标设备相关或相似的设备。
  10. 根据权利要求9所述设备故障预测方法，其特征在于，所述计算所述训练数据集中每条样本数据的权重，包括针对每台所述样本设备：
    将所述样本设备的特征数据与所述目标设备的特征数据进行区分；
    根据区分后的特征数据，训练分类模型；
    根据训练后的分类模型，计算所述样本设备每条特征数据的权重。
  11. 根据权利要求10所述设备故障预测方法，其特征在于，所述计算所述训练数据集中每条样本数据的权重，包括针对每台所述样本设备：
    将所述样本设备的特征数据标记为第一数据，将所述目标设备的特征数据标记为第二数据；
    根据所述第一数据和所述第二数据，训练分类模型，所述分类模型为基于联合学习的分类模型；
    根据训练后的分类模型，计算所述样本设备每条特征数据的权重，所述权重的计算公式为：
    Figure PCTCN2021101314-appb-100001
    其中，ω i为所述第一数据中第i条数据的权重，P 1i为所述第i条数据属于所述样本设备的概率，P 2i为所述第i条数据属于所述目标设备的概率。
  12. 根据权利要求9所述设备故障预测方法，其特征在于，所述利用所述权重训练得到目标设备的故障预测局部模型，包括：
    根据每台所述样本设备的特征数据、所述样本设备每条特征数据的权重和每台所述样本设备的故障数据，使用神经网络分别在所述样本设备有权重的所述训练数据集上训练得到所述目标设备的故障预测局部模型。
  13. 根据权利要求12所述设备故障预测方法，其特征在于，所述基于所述故障预测局部模型与联合学习算法，建立联合模型，包括：
    根据所述故障预测局部模型，使用联合学习算法反复迭代，获得所述样本设备在所述训练数据集上关于所述目标设备的联合模型。
  14. 根据权利要求9所述设备故障预测方法，其特征在于，所述目标设备和各个所述样本设备均为物联网中的边缘节点，所述目标设备的特征数据不对其他所述样本设备暴露，每个所述样本设备的特征数据和故障数据不对其他所述样本设备和所述目标设备暴露。
  15. 一种设备状态预测方法，其特征在于，包括：
    获取目标设备对应的至少两个特征数据、至少两个参考设备分别对应的至少两个模型训练数据以及各个所述模型训练数据分别对应的设备状态标签，所述目标设备和所述参考设备的设备种类相同；
    根据各个所述特征数据，确定所述模型训练数据对应的误差权重；
    根据各个所述模型训练数据、各个所述模型训练数据分别对应的设备状态标签以及各个所述模型训练数据分别对应的误差权重进行模型训练，以确定设备状态预测模型，所述设备状态预测模型用于所述目标设备的设备状态预测。
  16. 根据权利要求15所述的方法，其特征在于，所述根据各个所述特征数据，确定所述模型训练数据对应的误差权重，包括：
    从各个所述特征数据中确定出所述模型训练数据对应的替换数据；
    确定所述模型训练数据在其所属的参考设备对应的各个所述模型训练数据中的第一分布信息以及所述替换数据在各个所述特征数据中的第二分布信息；
    根据所述模型训练数据的第一分布信息以及所述模型训练数据对应的替换数据的第二分布信息，确定所述模型训练数据对应的误差权重。
  17. 根据权利要求16所述的方法，其特征在于，所述确定所述模型训练数据在其所属的参考设备对应的各个所述模型训练数据中的第一分布信息以及所述替换数据在各个所述特征数据中的第二分布信息，包括：
    基于所述模型训练数据在其所属的参考设备对应的第一非参数估计方法，确定所述模型训练数据在其所属的参考设备对应的各个所述模型训练数据中的第一分布概率，将所述第一分布概率确定为第一分布信息；
    基于第二非参数估计方法，确定所述替换数据在各个所述特征数据中的第二分布概率，将所述第二分布概率确定为第二分布信息。
  18. 根据权利要求17所述的方法，其特征在于，所述第一非参数估计方法和所述第二非参数估计方法均为核密度估计算法。
  19. 根据权利要求17所述的方法，其特征在于，所述根据所述模型训练数据的第一分布信息以及所述模型训练数据对应的替换数据的第二分布信息，确定所述模型训练数据对应的误差权重，包括：
    针对每个所述模型训练数据，确定所述模型训练数据对应的替换数据的第二分布概率和所述模型训练数据的第一分布概率的分布概率比值，将所述分布概率比值确定为确定所述模型训练数据对应的误差权重。
  20. 根据权利要求15所述的方法，其特征在于，所述方法应用到所述目标设备对应的边缘服务器或本地服务器上。
  21. 根据权利要求15所述的方法，其特征在于，所述根据各个所述模型训练数据、各个所述模型训练数据分别对应的设备状态标签以及各个所述模型训练数据分别对应的误差权重进行模型训练，以确定设备状态预测模型，包括：
    A1、根据将所述模型训练数据代入待预测模型中的预测结果以及所述模型训练数据对应的设备状态标签，确定所述模型训练数据对应的第一误差，根据所述模型训练数据对应的第一误差以及误差权重，以确定所述模型训练数据对应的第二误差；
    A2、判断是否满足迭代次数或者各个所述模型训练数据分别对应的第二误差是否满足预设条件，如果是，则将所述待预测模型确定为设备状态预测模型，如果否，则执行A3；
    A3、根据各个所述模型训练数据分别对应的第二误差，对所述待预测模型中的模型参数进行调整，以确定调整后的模型参数，并将所述待预测模型中的模型参数替换为所述调整后的模型参数，执行A1。
  22. 一种设备故障预测装置，其特征在于，包括：
    信息确定模块，用于确定目标设备的特征信息和目标设备对应的检测点数据信息；
    概率模型确定模块，用于基于所述目标设备的特征信息和目标设备对应的检测点数据信息，确定目标设备的特征数据概率分布模型和具有非共享数据的检测点数据的概率分布模型；
    权重确定模块，用于根据所述特征数据概率分布模型和所述检测点数据的概率分布模型，确定所述非共享数据的权重；
    模型建立模块，用于根据所述非共享数据、所述非共享数据的权重和所述非共享数据对应的设备故障标签，建立联邦学习模型；
    预测模块，用于根据所述联邦学习模型进行所述目标设备的设备故障预测。
  23. 一种设备故障预测装置，其特征在于，该设备故障预测装置包括：数据获取模块、权重计算模块、局部模型训练模块、联合模型建立模块和故障预测模块，其中，
    所述数据获取模块，用于根据目标设备属性，获取用于对目标设备建立预测模型的训练 数据集，其中所述数据集中的样本数据为共享数据；
    所述权重计算模块，用于计算所述训练数据集中每条样本数据的权重；
    所述局部模型训练模块，用于利用所述权重训练得到目标设备的故障预测局部模型；
    所述联合模型建立模块，用于基于所述故障预测局部模型与联合学习算法，建立联合模型；
    所述故障预测模块，用于根据所述联合模型对所述目标设备进行故障预测。
  24. 一种设备状态预测装置，其特征在于，包括：
    获取模块，用于获取目标设备对应的至少两个特征数据、至少两个参考设备分别对应的至少两个模型训练数据以及各个所述模型训练数据分别对应的设备状态标签，所述目标设备和所述参考设备的设备种类相同；
    权重确定模块，用于根据各个所述特征数据，确定所述模型训练数据对应的误差权重；
    训练模块，用于根据各个所述模型训练数据、各个所述模型训练数据分别对应的设备状态标签以及各个所述模型训练数据分别对应的误差权重进行模型训练，以确定设备状态预测模型，所述设备状态预测模型用于所述目标设备的设备状态预测。
  25. 一种可读介质，包括执行指令，当电子设备的处理器执行所述执行指令时，所述电子设备执行如权利要求1至21中任一所述的方法。
  26. 一种电子设备，包括处理器以及存储有执行指令的存储器，当所述处理器执行所述存储器存储的所述执行指令时，所述处理器执行如权利要求1至21中任一所述的方法。
PCT/CN2021/101314 2020-09-23 2021-06-21 预测方法、装置、可读介质及电子设备 Ceased WO2022062502A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
EP21870883.2A EP4131094A4 (en) 2020-09-23 2021-06-21 PREDICTION METHOD AND APPARATUS, READABLE MEDIUM AND ELECTRONIC DEVICE
JP2022564514A JP7783832B2 (ja) 2020-09-23 2021-06-21 予測方法、装置、可読媒体及び電子機器
US18/050,055 US20230070276A1 (en) 2020-09-23 2022-10-27 Prediction method and apparatus, readable medium and electronic device

Applications Claiming Priority (6)

Application Number Priority Date Filing Date Title
CN202011010325.X 2020-09-23
CN202011010347.6A CN114298362A (zh) 2020-09-23 2020-09-23 一种设备故障预测方法、装置、可读存储介质和计算设备
CN202011010325.XA CN114254790B (zh) 2020-09-23 2020-09-23 一种设备状态预测方法及装置
CN202011010347.6 2020-09-23
CN202011182449.6 2020-10-29
CN202011182449.6A CN114429235B (zh) 2020-10-29 2020-10-29 一种设备故障预测方法、装置、可读介质及电子设备

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/050,055 Continuation US20230070276A1 (en) 2020-09-23 2022-10-27 Prediction method and apparatus, readable medium and electronic device

Publications (1)

Publication Number Publication Date
WO2022062502A1 true WO2022062502A1 (zh) 2022-03-31

Family

ID=80844854

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/101314 Ceased WO2022062502A1 (zh) 2020-09-23 2021-06-21 预测方法、装置、可读介质及电子设备

Country Status (4)

Country Link
US (1) US20230070276A1 (zh)
EP (1) EP4131094A4 (zh)
JP (1) JP7783832B2 (zh)
WO (1) WO2022062502A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114818476A (zh) * 2022-04-01 2022-07-29 西南交通大学 一种应用于旋转机械设备寿命预测的联邦学习系统及方法
CN115169709A (zh) * 2022-07-18 2022-10-11 华能汕头海门发电有限责任公司 一种基于数据驱动的电站辅机故障诊断方法及系统
CN118037282A (zh) * 2024-04-15 2024-05-14 华中科技大学 基于ai和云技术的电力业务数据智能处理系统与方法
CN120161290A (zh) * 2025-04-22 2025-06-17 山东欢笑信息科技有限公司 一种面向电力系统的ai辅助短路故障定位系统及方法

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118298310B (zh) * 2024-04-19 2024-11-29 无锡市疾病预防控制中心 基于语义分割与集成学习的蚊虫分布预测方法和装置
CN118553408B (zh) * 2024-05-28 2024-11-15 深圳市雅士长华智能科技有限公司 数据处理方法、装置、设备及存储介质
CN118885374B (zh) * 2024-09-25 2024-12-17 杭州蜂鸟智能机器人有限公司 故障预测方法及计算机存储介质和终端设备
CN120493752B (zh) * 2025-05-22 2026-03-13 北京壹塔科技有限公司 温度场预测模型训练方法、温度场重建方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111158964A (zh) * 2019-11-26 2020-05-15 北京邮电大学 一种磁盘故障预测方法、系统、装置及存储介质
US20200210824A1 (en) * 2018-12-28 2020-07-02 Utopus Insights, Inc. Scalable system and method for forecasting wind turbine failure with varying lead time windows
CN111506862A (zh) * 2020-05-01 2020-08-07 西北工业大学 一种基于多源加权集成迁移学习的滚动轴承故障诊断方法
CN111678696A (zh) * 2020-06-17 2020-09-18 南昌航空大学 基于联邦学习的机械智能故障诊断方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2005111603A (ja) 2003-10-07 2005-04-28 Lazona Inc ロボットおよびロボット管理装置
JP2010092266A (ja) 2008-10-08 2010-04-22 Nec Corp 学習装置、学習方法及びプログラム
WO2018184224A1 (en) 2017-04-07 2018-10-11 Intel Corporation Methods and systems for boosting deep neural networks for deep learning
US10515295B2 (en) 2017-10-27 2019-12-24 Adobe Inc. Font recognition using triplet loss neural network training
JP2020047079A (ja) 2018-09-20 2020-03-26 富士通株式会社 学習プログラム、学習方法および学習装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200210824A1 (en) * 2018-12-28 2020-07-02 Utopus Insights, Inc. Scalable system and method for forecasting wind turbine failure with varying lead time windows
CN111158964A (zh) * 2019-11-26 2020-05-15 北京邮电大学 一种磁盘故障预测方法、系统、装置及存储介质
CN111506862A (zh) * 2020-05-01 2020-08-07 西北工业大学 一种基于多源加权集成迁移学习的滚动轴承故障诊断方法
CN111678696A (zh) * 2020-06-17 2020-09-18 南昌航空大学 基于联邦学习的机械智能故障诊断方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4131094A4 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114818476A (zh) * 2022-04-01 2022-07-29 西南交通大学 一种应用于旋转机械设备寿命预测的联邦学习系统及方法
CN114818476B (zh) * 2022-04-01 2023-08-22 西南交通大学 一种应用于旋转机械设备寿命预测的联邦学习系统及方法
CN115169709A (zh) * 2022-07-18 2022-10-11 华能汕头海门发电有限责任公司 一种基于数据驱动的电站辅机故障诊断方法及系统
CN115169709B (zh) * 2022-07-18 2023-04-18 华能汕头海门发电有限责任公司 一种基于数据驱动的电站辅机故障诊断方法及系统
CN118037282A (zh) * 2024-04-15 2024-05-14 华中科技大学 基于ai和云技术的电力业务数据智能处理系统与方法
CN120161290A (zh) * 2025-04-22 2025-06-17 山东欢笑信息科技有限公司 一种面向电力系统的ai辅助短路故障定位系统及方法

Also Published As

Publication number Publication date
JP7783832B2 (ja) 2025-12-10
US20230070276A1 (en) 2023-03-09
EP4131094A4 (en) 2023-11-15
EP4131094A1 (en) 2023-02-08
JP2023543100A (ja) 2023-10-13

Similar Documents

Publication Publication Date Title
WO2022062502A1 (zh) 预测方法、装置、可读介质及电子设备
CN114443398B (zh) 内存故障预测模型的生成方法、检测方法、装置及设备
CN108205424A (zh) 基于磁盘的数据迁移方法、装置及电子设备
CN114254792B (zh) 一种烟气含氧量预测方法及装置
WO2023115875A1 (zh) 硬件设备维护方法、装置及电子设备
CN115964943A (zh) 离线设备剩余寿命预测方法、装置和计算机设备
CN105183659B (zh) 基于多级模式预测的软件系统行为异常检测方法
CN120768750A (zh) 智能运维全局p2p架构下的故障自愈方法及系统
CN118838748A (zh) 内存故障的预测方法、装置、存储介质及电子设备
CN118802588A (zh) 丢包预测方法、装置、设备、介质及产品
CN111721542B (zh) 用于检测故障或模型失配的系统和方法
CN118838771A (zh) 一种容器容灾感知系统自适应基线预测及指标阈值调整方法
CN116170203A (zh) 一种安全风险事件的预测方法及系统
CN106681791A (zh) 一种基于对称邻居关系的增量式虚拟机异常检测方法
CN114429235A (zh) 一种设备故障预测方法、装置、可读介质及电子设备
CN120636011A (zh) 一种巡检方法、装置、设备、介质和计算机程序产品
CN120508421A (zh) 一种计算机系统故障的人工智能检测系统及方法
CN114298361A (zh) 一种设备状态信息的预测方法及装置
CN114363212A (zh) 一种设备检测方法、装置、设备和存储介质
CN119473741A (zh) 一种面向持久内存服务器的故障检测方法
CN114254790B (zh) 一种设备状态预测方法及装置
CN112764994B (zh) 一种容器云存储系统的灰色故障检测方法
CN115622264A (zh) 一种电力信息监控系统及方法
CN115550144A (zh) 分布式故障节点预测方法、装置、电子设备及存储介质
CN104080108B (zh) 一种针对无线传感网络数据的可变阈值异常点检测方法

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: 21870883

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2022564514

Country of ref document: JP

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 2021870883

Country of ref document: EP

Effective date: 20221028

NENP Non-entry into the national phase

Ref country code: DE