CN117874596A

CN117874596A - Method and system for evaluating health state of equipment with multitask learning and privacy protection

Info

Publication number: CN117874596A
Application number: CN202311706679.1A
Authority: CN
Inventors: 张永; 尹晓宇; 周炜; 郑英; 严保康
Original assignee: Wuhan University of Science and Engineering WUSE
Current assignee: Wuhan University of Science and Engineering WUSE
Priority date: 2023-12-11
Filing date: 2023-12-11
Publication date: 2024-04-12

Abstract

The invention provides a method and a system for evaluating the health state of equipment with multitask learning and privacy protection, wherein the method comprises the following steps: singular value decomposition is carried out on historical operation parameters of each sample device in the first client to obtain time domain features and frequency domain features, and then feature reconstruction and mutation detection are carried out to obtain multi-task labels of each sample device; constructing a source domain training set based on the multi-task tag, the time domain feature and the frequency domain feature of the first sample device in the source domain, and constructing a target domain training set based on the multi-task tag, the time domain feature and the frequency domain feature of the second sample device in the target domain; based on the source domain training set and the target domain training set in the first client, and the source domain training set and the target domain training set in each second client, performing federal learning training of countermeasure domain self-adaption, obtaining a target domain multitask evaluation model, performing equipment health state evaluation, and improving the accuracy of equipment health state evaluation while protecting data privacy.

Description

Method and system for evaluating health state of equipment with multitask learning and privacy protection

Technical Field

The invention relates to the technical field of artificial intelligence, in particular to a method and a system for evaluating the health state of equipment with multitask learning and privacy protection.

Background

Rotating equipment plays an important role in modern industrial equipment. They are an integral part and are widely used in various fields. These devices, in their unique construction and function, play a critical role in the aerospace, transportation, electrical, petroleum industries, and the like. The rotating equipment consists of a rotor, a gear, a bearing and other parts, and the high-efficiency operation is realized through the power transmission unit. However, long-term use causes fatigue and wear, and therefore maintenance must be regularly performed to ensure proper operation of the apparatus. If maintenance work is ignored, not only the performance and the service life of equipment can be affected, but also serious property loss and even personal safety risks can be brought. Therefore, the equipment is accurately assessed for health status, and a reasonable maintenance scheme is formulated according to the equipment, so that the safety of the equipment is ensured, and the maintenance cost is saved.

The existing deep learning algorithm has a good prediction effect when the training set and the testing set have the same distribution condition, however, in actual production, different devices have different operation parameters, and in addition, due to the consideration of safety and the like, enterprises cannot share own data to carry out privacy protection, a data island is formed, so that the data acquired by actual testing application and the data acquired by training are difficult to achieve consistency, and the health state of the devices is difficult to accurately evaluate.

Therefore, there is a need for a device health status assessment method and system for multitasking learning and privacy protection to solve the above problems.

Disclosure of Invention

The invention provides a method and a system for evaluating the health state of equipment for multitask learning and privacy protection, which are used for solving the defect that the data acquired by actual test application and the data acquired by training in the prior art are inconsistent in distribution and difficult to evaluate the health state of the equipment accurately, and improving the accuracy of the health state evaluation of the equipment while protecting the privacy of the data.

The invention provides a device health state assessment method for multitasking learning and privacy protection, which is applied to a first client side in a plurality of client sides and comprises the following steps:

singular value decomposition is carried out on historical operation parameters of each sample device in the first client to obtain time domain features and frequency domain features corresponding to the historical operation parameters; the sample devices comprise a first sample device under a source domain and a second sample device under a target domain;

performing feature reconstruction on the time domain features and the frequency domain features to obtain health indexes of the sample devices;

carrying out mutation inspection on the health index, and acquiring a multi-task label of each sample device according to mutation inspection results; the multi-task labels comprise a residual life label, a reliability evaluation label and a health stage division label;

Constructing a source domain training set based on the multi-task tag, the time domain feature and the frequency domain feature of the first sample device in the source domain, and constructing a target domain training set based on the multi-task tag, the time domain feature and the frequency domain feature of the second sample device in the target domain;

performing adaptive federal learning training in the countermeasure domain based on a source domain training set and a target domain training set in the first client and a source domain training set and a target domain training set in each second client except the first client in a plurality of clients to obtain a target domain multitask evaluation model;

the target domain multitask evaluation model is used for outputting a residual life predicted value, a reliability evaluation predicted value and a health stage division predicted value of target equipment according to time domain features and frequency domain features of the target equipment in the target domain.

According to the method for evaluating the health state of the equipment for multi-task learning and privacy protection provided by the invention, the singular value decomposition is carried out on the historical operation parameters of each sample equipment in the first client to obtain the time domain characteristics and the frequency domain characteristics corresponding to the historical operation parameters, and the method comprises the following steps:

For each of the sample devices, performing the following operations:

constructing a time domain parameter vector of the current sample equipment according to historical operation parameters acquired by each sampling point in the whole life operation period of the current sample equipment; the first dimension of the time domain parameter vector represents different operation periods, the second dimension represents different sampling points, and the third dimension represents historical operation parameters collected at each sampling point;

averaging historical operation parameters corresponding to the third dimension of the time domain parameter vector to obtain a time domain average value of the time domain parameter vector;

calculating the difference between the time domain parameter vector and the time domain average value to obtain a time domain average tensor;

performing tensor singular value decomposition on the time-domain average tensor to obtain a time-domain orthogonal vector;

performing Fourier transform on the historical operation parameters corresponding to the third dimension of the time domain average tensor and the historical operation parameters corresponding to the third dimension of the time domain orthogonal vector respectively to obtain a frequency domain average tensor and a frequency domain orthogonal vector;

performing dimension reduction processing on a plurality of sampling points corresponding to the second dimension of the time domain orthogonal vector and a plurality of sampling points corresponding to the second dimension of the frequency domain orthogonal vector;

And acquiring the time domain characteristics according to the time domain orthogonal vector after the dimension reduction and the time domain average tensor, and acquiring the frequency domain characteristics according to the frequency domain orthogonal vector after the dimension reduction and the frequency domain average tensor.

According to the method for evaluating the health state of the equipment for multi-task learning and privacy protection provided by the invention, the feature reconstruction is carried out on the time domain features and the frequency domain features to obtain the health index of each sample equipment, and the method comprises the following steps:

inputting the time domain features, the frequency domain features and randomly generated noise data into a coding network of a target self-encoder, and reconstructing to obtain the health index;

the target self-encoder is obtained by performing unsupervised training on the self-encoder based on a long-short-time memory network based on time domain features and frequency domain features of each sample device.

According to the method for evaluating the health state of the equipment for multi-task learning and privacy protection provided by the invention, mutation detection is carried out on the health index, and the multi-task label of each sample equipment is obtained according to the mutation detection result, and the method comprises the following steps:

carrying out mutation inspection on the health index to obtain variation trend information of the health index;

Respectively carrying out health stage division marking and residual life marking on the health index according to the change trend information to obtain the residual life label and the health stage division label;

inputting time domain features and frequency domain features corresponding to the historical operation parameters of each sample device to a reliability evaluation model to obtain the reliability evaluation label;

the reliability evaluation model is obtained by carrying out maximum likelihood estimation on a logistic regression model based on a health stage division label of any one sample device and time domain features and frequency domain features corresponding to historical operation parameters of any one sample device.

According to the equipment health state evaluation method for multi-task learning and privacy protection provided by the invention, the target domain multi-task evaluation model comprises a target domain coding model and a target domain multi-task network;

the target domain multitasking network comprises a multi-expert mixed model, a residual life prediction task branch network, a reliability evaluation prediction task branch network, a health phase division prediction task branch network, a first self-attention network and a second self-attention network;

the health stage division prediction task branch network comprises a first characteristic tower layer, a first characteristic fusion layer and a first prediction layer, the residual life prediction task branch network comprises a second characteristic tower layer, a second characteristic fusion layer and a second prediction layer, and the reliability evaluation prediction task branch network comprises a third characteristic tower layer, a third characteristic fusion layer and a third prediction layer;

The output end of the target domain coding model is connected with the input end of the multi-expert mixed model; the output end of the multi-expert hybrid model is respectively connected with the input end of the first characteristic tower layer, the input end of the second characteristic tower layer and the input end of the third characteristic tower layer;

the output end of the first characteristic tower layer is respectively connected with the input end of the first characteristic fusion layer and the input end of the first self-attention network; the output end of the first characteristic fusion layer is connected with the input end of the first prediction layer;

the output end of the first self-attention network and the output end of the second characteristic tower layer are connected with the input end of the second characteristic fusion layer; the output end of the second characteristic fusion layer is connected with the input end of the second prediction layer;

the output end of the second characteristic fusion layer and the output end of the second characteristic tower layer are connected with the input end of the second self-attention network; the output end of the second self-attention network and the output end of the third characteristic tower layer are connected with the input end of the third characteristic fusion layer; and the output end of the third characteristic fusion layer is connected with the input end of the third prediction layer.

According to the method for evaluating the health state of the equipment for multi-task learning and privacy protection provided by the invention, the method for evaluating the health state of the equipment for multi-task learning and privacy protection performs adaptive federal learning training in the countermeasure field based on a source domain training set and a target domain training set in the first client and a source domain training set and a target domain training set in each second client except the first client in a plurality of clients, and obtains a target domain multi-task evaluation model, and comprises the following steps:

performing federal learning training on a source domain multi-task assessment model in the first client based on multi-task labels corresponding to each first sample device in a source domain training set in the first client and multi-task labels corresponding to each first sample device in a source domain training set in each second client;

constructing an initial target domain coding model according to a source domain coding model in the trained source domain multitasking evaluation model, and constructing an initial target domain multitasking model according to a source domain multitasking network in the trained source domain multitasking evaluation model;

performing federal learning training of countermeasure domain adaptation on an initial target domain coding model in the first client based on a domain label corresponding to each first sample device in a source domain training set and a domain label corresponding to each second sample device in a target domain training set in the first client, and a domain label corresponding to each first sample device in the source domain training set and a domain label corresponding to each second sample device in the target domain training set in each second client, so as to obtain the target domain coding model;

And performing federal learning training on the initial target domain multi-task model based on the target domain coding model, the multi-task labels corresponding to the second sample devices in the target domain training set in the first client and the multi-task labels corresponding to the second sample devices in the target domain training set in the second client to obtain the target domain multi-task model.

According to the equipment health state assessment method for multi-task learning and privacy protection provided by the invention, the loss function corresponding to the initial target domain coding model in the first client is constructed jointly based on the correctness probability value of the domain prediction result corresponding to each first sample equipment in the source domain training set in the first client and the correctness probability value of the domain prediction result corresponding to each second sample equipment in the target domain training set.

The invention also provides a system for evaluating the health state of the equipment for multitasking learning and privacy protection, which is applied to any first client side of a plurality of client sides and comprises the following steps:

the first extraction module is used for carrying out singular value decomposition on the historical operation parameters of each sample device in the first client to obtain time domain features and frequency domain features corresponding to the historical operation parameters; the sample devices comprise a first sample device under a source domain and a second sample device under a target domain;

The second extraction module is used for carrying out feature reconstruction on the time domain features and the frequency domain features to obtain health indexes of the sample equipment;

the marking module is used for carrying out mutation detection on the health indexes and acquiring the multi-task labels of the sample equipment according to mutation detection results; the multi-task labels comprise a residual life label, a reliability evaluation label and a health stage division label;

the construction module is used for constructing a source domain training set based on the multi-task label, the time domain feature and the frequency domain feature of the first sample equipment in the source domain and constructing a target domain training set based on the multi-task label, the time domain feature and the frequency domain feature of the second sample equipment in the target domain;

the training module is used for performing self-adaptive federal learning training in the countermeasure field based on a source domain training set and a target domain training set in the first client and the source domain training set and the target domain training set in each second client except the first client in a plurality of clients to obtain a target domain multitask evaluation model;

The invention also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the device health state assessment method for multi-task learning and privacy protection as described above when executing the program.

The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements a device health status assessment method for multitasking learning and privacy protection as described in any of the above.

The invention also provides a computer program product comprising a computer program which when executed by a processor implements a device health state assessment method for multitasking learning and privacy protection as described in any of the above.

According to the equipment health state assessment method and system for multi-task learning and privacy protection, on one hand, the training data set is constructed efficiently and accurately through singular value decomposition and mutation inspection, and the accuracy and efficiency of training of a subsequent target domain multi-task assessment model are improved; on the other hand, the field self-adaptive training is carried out through the training data set to realize the feature mapping and migration of the source domain and the target domain, and the federal learning framework with stronger robustness is introduced to carry out the collaborative training of the target domain state evaluation model, so that the alignment of the source domain data and the target domain data is realized, the data privacy is protected, the expansion of the data set is realized, and the accuracy, the generalization and the robustness of the equipment health state evaluation are improved.

Drawings

In order to more clearly illustrate the invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow diagram of a method for evaluating the health status of a device for multi-task learning and privacy protection provided by the present invention;

FIG. 2 is a second flow chart of a method for evaluating health status of a device for multi-task learning and privacy protection according to the present invention;

FIG. 3 is a graphical representation of health indicators provided by the present invention;

FIG. 4 is a schematic diagram of the distribution of reliability and health staging provided by the present invention;

FIG. 5 is a schematic diagram of a target domain multitasking evaluation model provided by the present invention;

FIG. 6 is a schematic error diagram of a reliability assessment provided by the present invention;

FIG. 7 is a schematic error diagram of a residual life prediction provided by the present invention;

FIG. 8 is a schematic diagram of a simulation of the health staging provided by the present invention;

FIG. 9 is a schematic diagram of a device health status assessment system for multi-task learning and privacy protection provided by the present invention;

Fig. 10 is a schematic structural diagram of an electronic device provided by the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

The health status assessment of the device mainly comprises health staging, life prediction and reliability assessment. Because of the complex operating conditions of the equipment, how to accurately evaluate the health status of the equipment is one of the difficulties to be solved. The existing deep learning algorithm has good prediction effect when the training set and the test set have the same distribution and the training data set is sufficient, however, different devices have different operation parameters in the actual production process, and the data collected by one factory in the actual production process is difficult to reflect the condition of the full life cycle of the bearing, so that the data volume can be expanded by obtaining the data of other factories. However, due to the consideration of safety and the like among different enterprises, the enterprises cannot share own data to carry out privacy protection, and form a data island, so that the data acquired by the actual test application and the data acquired by training are difficult to achieve consistency, and the health state of the equipment is difficult to accurately evaluate.

Aiming at the problems, the embodiment provides a method and a system for evaluating the health state of equipment by multi-task learning and privacy protection, wherein the health state evaluation of the equipment is realized by combining self-adaptive learning, federal learning and multi-task learning in the countermeasure field, firstly, the alignment between source domain and target domain data is realized by transfer learning, the federal learning is introduced, the privacy protection is realized, and finally, three tasks of equipment life prediction, health stage division and reliability evaluation are synchronously predicted by multi-task learning to obtain the health state evaluation result of the equipment, thereby realizing the efficient and accurate health state evaluation.

The device health status assessment method of the present invention for multitasking learning and privacy protection is described below in conjunction with fig. 1-8. The method can be applied to a federal learning framework comprising a plurality of clients, and the specific number of the clients contained in the federal learning framework can be set according to actual requirements; the first client is any client in the decentralized federal learning framework, and the second client is other clients in the federal learning framework except the first client. The method is executed as a first client in the federal learning framework. As shown in fig. 1, the method includes:

Step 101, performing singular value decomposition on historical operation parameters of each sample device in the first client to obtain time domain features and frequency domain features corresponding to the historical operation parameters; the sample devices comprise a first sample device under a source domain and a second sample device under a target domain;

it should be noted that the method for evaluating the health status of the equipment may be applicable to various equipment for evaluating health status, including but not limited to various mechanical equipment in fans, helicopters, engines, etc.; the mechanical device may be a bearing or other device.

The historical operation parameters are parameters for representing the historical operation state of the sample equipment, and can be determined according to the actual operation scene of the equipment; for example, the historical operating parameters of the rolling bearing are vibration signals in x-axis and y-axis directions and temperature signals acquired at a preset frequency in each historical operating period, and 2560 points can be acquired in the operating period.

It should be noted that, in the federal learning framework, the local memory of the first client and the local memory of each second client store the historical operation parameters of the first sample device under the active domain and the second sample device under the target domain, and the source domain multitask evaluation model; in the process of training to obtain the target domain multitask evaluation model, the first client and each second client participate in the federal learning training of the anti-domain self-adaption. The target domain is the scene domain under the model application; the source domain is the scene domain under model training.

As shown in FIG. 2, X _S For source domain data sets, MT _S X for source domain multitasking assessment model _T Is the target domain dataset. Because the historical operation parameters acquired by the sensors often contain a large amount of noise data and have higher data dimension, if the model training effect is directly carried out, the subsequent data analysis and model training are not facilitated. In order to solve the problem, the historical operation parameters of each sample device in the first client are subjected to feature extraction by using tensor singular value decomposition so as to extract time domain features and frequency domain features, thereby realizing the elimination of data reduction and redundant information.

The time domain feature and frequency domain feature modes can be that SVD (Singular Value Decomposition ) is carried out on historical operation parameter data to obtain a singular value matrix and left and right singular vectors, and then the singular value matrix is directly used as time domain feature, or after screening or dimension reduction is carried out on the singular value matrix, the singular value matrix is used as time domain feature; and converting the time domain features to the frequency domain by performing Fourier transform by using the left singular vector and the right singular vector.

102, performing feature reconstruction on the time domain features and the frequency domain features to obtain health indexes of the sample devices;

Optionally, after the time domain feature and the frequency domain feature are obtained, the time domain feature and the frequency domain feature are three-dimensional, so that feature reconstruction can be performed on the time domain feature and the frequency domain feature, so that the time domain feature and the frequency domain feature are fused and reduced to two-dimensional data, health indexes of various sample devices are obtained, the running state and the performance condition of the devices are better reflected, the label marking is conveniently and rapidly performed, and the accuracy and the efficiency of the health state assessment of the devices are further improved.

Wherein the feature reconstruction may be based on generation of a countermeasure network or a self-encoder implementation, which is not specifically limited in this embodiment.

Step 103, carrying out mutation detection on the health index, and obtaining a multi-task label of each sample device according to a mutation detection result; the multi-task labels comprise a residual life label, a reliability evaluation label and a health stage division label;

optionally, after the health index is obtained, mutation inspection may be performed on the health index to obtain variation trend information of the health index, so as to perform marking of the multi-task label based on the variation trend information, thereby obtaining a remaining life label, a reliability evaluation label and a health stage division label.

104, constructing a source domain training set based on the multi-task tag, the time domain feature and the frequency domain feature of the first sample device in the source domain, and constructing a target domain training set based on the multi-task tag, the time domain feature and the frequency domain feature of the second sample device in the target domain;

as shown in fig. 2, a source domain training set is constructed by taking time domain features and frequency domain features of a first sample device under a source domain as samples and taking a multi-task label and a domain label of the first sample device as labels; and similarly, taking the time domain characteristics and the frequency domain characteristics of the second sample equipment in the target domain as samples, and taking the multi-task labels and the domain labels of the second sample equipment as labels to construct the target training set.

It should be noted that, for each client, the processing of steps 101 to 104 may be performed on the historical operating parameters of the local sample device according to the above method to obtain the local source domain training set and the target domain training set, which are not described herein again.

Step 105, performing adaptive federal learning training in the countermeasure domain based on the source domain training set and the target domain training set in the first client and the source domain training set and the target domain training set in each second client except the first client in a plurality of clients to obtain a target domain multitask evaluation model; the target domain multitask evaluation model is used for outputting a residual life predicted value, a reliability evaluation predicted value and a health stage division predicted value of target equipment according to time domain features and frequency domain features of the target equipment in the target domain.

The target domain multitasking evaluation model may be a target domain multitasking network including a target domain coding model; the target domain coding model can be constructed and generated based on one or more neural networks, such as a convolution network, a full-connection network and the like, and is used for extracting coding characteristics of time domain characteristics and frequency domain characteristics of equipment; the target domain multiplexing network may also be generated based on one or more neural network constructions, such as a convolutional network, a fully-connected network, an attention network, etc., and is used for performing multiplexing learning according to the coding features output by the target domain coding model, so as to output corresponding residual life prediction values, reliability evaluation prediction values and health stage division prediction values.

Optionally, the adaptive federal learning training of the challenge domain may be performed based on the source domain training set and the target domain training set in the first client, and the source domain training set and the target domain training set in each second client, to obtain a target domain multitask evaluation model.

The training mode of the target domain multitask evaluation model can be that the source domain multitask evaluation model in the first client is subjected to federal learning training based on a source domain training set in the first client and a source domain training set in each second client; then, determining an initial target domain coding model and an initial target domain multitasking network according to the trained source domain multitasking evaluation model; performing adaptive federal learning training of the countermeasure field on the initial target domain coding model based on the source domain training set and the target domain training set in the first client and the source domain training set and the target domain training set in each second client to obtain a target domain coding model; directly taking the initial target domain multi-task network as a final target domain multi-task network; or performing fine tuning training on the initial target domain multi-task network based on the target domain training set in the first client and the target domain training set in each second client to obtain a final target domain multi-task network.

Then, after the current operation parameter of the target device in the target domain is received in the first client, singular value decomposition may be performed on the current operation parameter by referring to step 101, so as to obtain a time domain feature and a frequency domain feature corresponding to the current operation parameter; and then, inputting the time domain features and the frequency domain features corresponding to the current operation parameters into the target domain multitask evaluation model trained in the step 105, thereby obtaining the residual life predicted value, the reliability evaluation predicted value and the health stage division predicted value of the target equipment.

According to the equipment health state assessment method for multi-task learning and privacy protection, on one hand, the training data set is constructed efficiently and accurately through singular value decomposition and mutation inspection, and the accuracy and the efficiency of training a subsequent target domain multi-task assessment model are improved; on the other hand, the field self-adaptive training is carried out through the training data set to realize the feature mapping and migration of the source domain and the target domain, and the federal learning framework with stronger robustness is introduced to carry out the collaborative training of the target domain state evaluation model, so that the alignment of the source domain data and the target domain data is realized, the data privacy is protected, the expansion of the data set is realized, and the accuracy, the generalization and the robustness of the equipment health state evaluation are improved.

In some embodiments, performing singular value decomposition on the historical operating parameters of each sample device in the first client to obtain a time domain feature and a frequency domain feature corresponding to the historical operating parameters, where the singular value decomposition includes:

for each of the sample devices, performing the following operations:

Optionally, the step of extracting the time domain features and the frequency domain features in step 101 includes:

the following is performed for each sample device:

firstly, constructing a time domain parameter vector of current sample equipment according to historical operation parameters acquired by each sampling point in the whole life operation period of the current sample equipmentI.e. the first dimension I of the time domain parameter vector O of the current sample device ₁ For different transportationLine period, second dimension I ₂ For different sampling points, a third dimension I ₃ Historical operating parameters collected at each sampling point; for example, in the case that the sample device is a bearing device, the number of life-time operation cycles of the bearing device is 2375, the number of sampling points in each operation cycle is 2560, the historical operation parameters collected at each sampling point are vibration signals in x-axis and y-axis directions and temperature signals, and then the vector dimension of the time-domain parameter vector O of the device is 2375×2560×3.

Then, calculating the average value of historical operation parameters of the time domain parameter vector O along the third dimension to obtain a time domain average value M of the time domain parameter vector O; then, calculating the difference between the time domain parameter vector O and the time domain average value M to obtain a time domain average tensor A, namely A=O-M; then, tensor singular value decomposition is carried out on the time domain average tensor A to obtain a time domain orthogonal vectorAnd performing Fourier transform along the third dimension of the time domain average tensor A and the third dimension of the time domain orthogonal vector U to obtain a frequency domain average tensor ∈ ->And frequency domain orthogonal vector->

Next, to reduce redundant information and noise data in the data, orthogonal vectors in the time domainOrthogonal vector +.>Data truncation is carried out or a dimension reduction algorithm is adopted in the second dimension of the (2), such as principal component analysis, so as to obtain a dimension-reduced time domain orthogonal vector ++>And the frequency domain orthogonal vector after dimension reduction

Then, the time domain orthogonal vector after the dimension reductionPerforming matrix star multiplication operation with the time domain average tensor A to obtain time domain characteristic T, namely time domain characteristic +.>For the frequency domain orthogonal vector after dimension reduction +.>Mean tensor from frequency domain->Performing matrix dot multiplication to obtain frequency domain feature F, namely frequency domain feature +. >Where represents the matrix star multiplier, represents the matrix dot multiplier.

The method provided by the embodiment can extract important characteristics of the equipment from the historical operation parameters by combining singular value decomposition and dimension reduction, so that the performance and training efficiency of the target domain multitask evaluation model are improved.

In some embodiments, the performing feature reconstruction on the time domain feature and the frequency domain feature to obtain a health indicator of each sample device includes:

Optionally, the step of obtaining the health indicator in step 102 includes:

because the time domain features and the frequency domain features obtained through the step 101 are three-dimensional data, in order to better reflect the running state and the performance condition of the equipment, the label is marked more quickly and accurately, and the time domain features, the frequency domain features and the randomly generated noise data are reconstructed by using the coding network of the target self-encoder to obtain corresponding health indexes; as shown in fig. 3, the operation state and the performance condition of the whole life cycle of the sample equipment can be effectively represented by reconstructing to obtain the corresponding health index.

The target self-encoder is a noise reduction self-encoder constructed based on Long Short-Term Memory (LSTM), hereinafter also referred to as LSTM noise reduction self-encoder. The LSTM de-noising self-encoder is used to reconstruct the input data from noisy data, specifically by adding noise to the input data and mapping the noisy data to a potential representation through the encoding network, and finally reconstructing the input data through the decoder.

According to the method provided by the embodiment, the LSTM denoising self-encoder is used for carrying out feature fusion and reconstruction of the time domain features and the frequency domain features, so that the influence of noise in the historical operation parameters of the sample equipment on subsequent data analysis can be effectively reduced.

In some embodiments, the performing mutation test on the health indicator, and obtaining the multi-task label of each sample device according to the mutation test result, includes:

Optionally, the step of constructing the multi-tasking label in step 103 includes: firstly, the mutation detection algorithm is used for classifying the health index to obtain the health classification label. And secondly, constructing a reliability label by using logistic regression, wherein the specific flow is as follows:

firstly, mutation detection is carried out on the health index by using a mutation detection algorithm to obtain the change trend information of the health index, wherein the change trend information of the health index comprises monotone trends in time series data of the health index, such as ascending, descending or no trend and mutation points.

The mutation detection algorithm has good statistical performance in detecting data mutation points with different distributions, so that the health state can be adaptively divided by using the method. Mutation testing is a commonly used non-parametric statistical method for detecting monotonic trends (rising, falling, or no trend) in time series data. The main principle is to compare each data point in the time series with all the data points before it. For each data point, the number of data points smaller than it and the number of data points larger than it are calculated, and then the magnitude relationship of these two numbers is compared. If the number of data points smaller than it is significantly greater than the number of data points larger than it, then a downward trend is indicated; conversely, if the number of data points greater than it is significantly greater than the number of data points less than it, an upward trend is indicated. If the two numbers are equal, then there is no trend. The so-called mutation points are determined based on the following steps:

For health index, data at a certain time and before are taken to form a new time sequence X= { X ₁ ,x ₂ ,…,x _n Sequence of construction XWhere j=1, 2, …, i; k=1, 2, …, n. At time of dayUnder the assumption that the inter-sequence X is independent randomly, S is obtained _k The mean and variance of (a) are: e [ S ] _k ]＝k(k-1)/4，var[S _k ]=k (k-1) (2k+5)/72, where 1.ltoreq.k.ltoreq.n, S _k Standardization: />Wherein UF is ₁ ＝0，UF _i Is standard normal distribution, which is a statistic sequence calculated according to time sequence X, given significance level alpha, and the normal distribution table is checked to obtain a critical value U corresponding to the significance level _α The method comprises the steps of carrying out a first treatment on the surface of the If UF is _i |>U _α Then a significant trend in the sequence is indicated. Then UB is obtained according to the reverse sequence of X _k If UF _k And UB _k And the intersection points of the two curves are in the range of the two critical lines, and the intersection points are abrupt points. Next, the point after the point of mutation is intercepted to construct a new time sequence, and the steps are repeated until all data detection is completed.

As shown in fig. 4, after acquiring the monotone trend and the mutation point in the health indicator time series data, the number of stages to be divided may be determined based on the monotone trend, and the demarcation points of different stages may be determined based on the mutation point, so as to divide the whole life cycle of the sample device into stage 0, stage 1, stage 2 and stage 3; wherein, stage 0 is the initial stage, stage 1 is the small-amplitude degradation stage, stage 2 is the moderate degradation stage, and stage 3 is the failure stage.

Next, determining a remaining life marker based on the health staging markers; the remaining life markers are obtained from the point calculation where the device starts to degrade; i.e. before the degradation point, the residual life prediction actual value defaults to 100%, and after the degradation point, the residual life actual value of the sample device is calculated and obtained according to the total life cycle of the sample device, the time period of the degradation point and the current period.

As shown in fig. 4, the first sampling point of the phase 2 may be set as a degradation point, and the residual life flag of the bearing is calculated from the degradation point, thereby more conforming to the actual scenario.

And then, inputting the time domain features and the frequency domain features corresponding to the historical operation parameters of each sample device into a health stage division label based on any sample device, and carrying out maximum likelihood estimation on the logistic regression model to obtain a reliability evaluation model according to the time domain features and the frequency domain features corresponding to the historical operation parameters of the sample device, so as to obtain the reliability evaluation label by fitting.

The so-called logistic regression model is a binary classification model that depends on dependent variables. In particular, the probabilities associated with different health phases of the machine are calculated from a set of characteristic parameters. For example, use X _i (t)＝(x ₁ (t),x ₂ (t)···x _i (t)) to represent a feature set formed by the time domain features and the frequency domain features of the sample device. According to the health stage division label, determining an output label of the model, wherein the state is expressed as y (t) =1 when the equipment operates normally at the time t, and the bearing fault state is expressed as y (t) =0 when the equipment fails to operate at the time t. In x according to definition of reliability evaluation _i As the input quantity of the logistic regression model, outputting a label by a model determined by a health stage division label, and obtaining the regression coefficient of the logistic regression model through maximum likelihood estimation, thereby constructing and obtaining a reliability evaluation model; the calculation formula of the reliability evaluation model is as follows:

wherein R (t|X _i (t)) is a reliability evaluation tag for the run period t, { k ₀ ,k ₁ ,…,k _i And the regression coefficient of the reliability evaluation model.

As shown in fig. 4, the change from 1 to 0 of the device reliability evaluation tag reflects the degree of degradation and failure thereof. In general, the closer the reliability evaluation tag is to 1, the higher the reliability of the device is, i.e. the performance and service life of the device in the working process are in a better state; and the closer the reliability evaluation tag is to 0, the lower the reliability of the device, i.e. the device is approaching or has failed, and normal operation cannot be continued.

According to the method, the device and the system, the multi-task labels can be determined efficiently and accurately through the mutation detection and logistic regression model, so that training efficiency and evaluation performance of the multi-task evaluation model are improved.

In some embodiments, the target domain multitasking evaluation model includes a target domain coding model and a target domain multitasking network;

As shown in fig. 5, the target domain multitasking evaluation model includes a target domain coding model and a target domain multitasking network; the target domain coding model can be constructed and generated by a convolution coder and other coding models such as a derivative or improved coding model thereof.

the target domain coding model is used for extracting coding features of time domain features and frequency domain features of target equipment, and the multi-expert mixed model is used for extracting mixed features of the multi-expert model for the coding features. The first feature tower layer is used for carrying out multi-scale extraction on the mixed features, and the first feature fusion layer is used for fusing the multi-scale features output by the first feature tower layer; the first prediction layer is used for predicting and outputting a health stage division predicted value according to the fusion characteristics output by the first characteristic fusion layer.

The second feature tower layer is used for carrying out multi-scale extraction on the mixed features, the first self-attention network is used for carrying out attention learning on the fused features output by the first feature fusion layer, and the second feature fusion layer is used for fusing the multi-scale features output by the second feature tower layer and the attention features output by the first self-attention network; the second prediction layer is used for predicting and outputting a residual life prediction value according to the fusion characteristics output by the second characteristic fusion layer.

The third feature tower layer is used for carrying out multi-scale extraction on the mixed features, the second self-attention network is used for carrying out attention learning on the fusion features output by the second feature fusion layer and the multi-scale output by the second feature tower layer; the third feature fusion layer is used for fusing the multiscale features output by the third feature tower layer and the attention features output by the second self-attention network; the third prediction layer is used for predicting and outputting a reliability evaluation predicted value according to the fusion characteristics output by the third characteristic fusion layer.

In order to keep information balance among a plurality of tasks and avoid potential output mismatch, the method provided by the embodiment introduces the output of each task branch network into an adjacent task branch network through a self-attention mechanism so as to establish an effective information flow channel between a classification task (namely a health stage division task) and other two regression tasks (namely a life prediction task and a reliability evaluation task), promote information exchange among the multitasks, improve the accuracy of life prediction through health stage division, provide the accuracy of reliability evaluation through life prediction results, realize the relation drive among the plurality of tasks, ensure the consistency of the output, realize good balance state in various tasks, ensure synchronous output of multiple target values, avoid potential output mismatch and further improve the accuracy of equipment state evaluation.

In some embodiments, the performing the federal learning training of the challenge domain adaptation based on the source domain training set and the target domain training set in the first client and the source domain training set and the target domain training set in each of the second clients except the first client to obtain a target domain multitasking evaluation model includes:

It should be noted that, in the federal learning process, the source domain multitask evaluation model in each client is trained locally, and when the group synchronization condition is satisfied, local parameters of each model are combined by randomly selecting a leader (i.e., the first client) to realize parameter update of the model, thereby solving the data privacy problem and realizing the multitask prediction function. In the training process, each client independently builds a source domain multitask evaluation model on a private source domain training set and carries out local model training, and when all the models are trained, the leaders can combine local parameters from all the models by dynamically selecting the leaders so as to realize global updating of the models. Finally, the leader transmits the global parameters to other models, so that each model is updated. Through the training process, the training of the global model can be realized under the condition of protecting the data privacy.

As shown in fig. 2, the training step of the target domain multitasking model in step 105 specifically includes:

firstly, performing federal learning training on a source domain multitask evaluation model, wherein the method comprises the following specific steps of:

for current model training, the first client communicates with each second client to receive local model parameters obtained by carrying out local iterative training on each locally stored source domain multitask evaluation model based on each local source domain training set and global model parameters of the source domain multitask evaluation model corresponding to the last model training in the current model training process of each second client. In the process that each client trains the source domain multi-task assessment model locally, iterative training is performed on the source domain multi-task assessment model by taking the minimum loss function constructed by the multi-task labels and the multi-task predicted values of the sample equipment as a target.

The first client can aggregate and update the locally stored source domain multi-task assessment model according to the received local model parameters of the source domain multi-task assessment model and the locally stored source domain training set to obtain global model parameters of the source domain multi-task assessment model corresponding to the current model training.

After the first client acquires the global model parameters of the source domain multi-task assessment model corresponding to the current model training, the first client sends the global model parameters corresponding to the current model training to each second client, and the model training step is executed iteratively until the source domain multi-task assessment model corresponding to the global model parameters meets the preset termination condition; the preset termination conditions include, but are not limited to, the number of model training times being greater than the preset number of times, model convergence, etc., which are not particularly limited in this embodiment.

And training global model parameters of the corresponding source domain multitask evaluation model according to the last model to obtain the trained source domain multitask evaluation model.

In some embodiments, for each client, the loss function of the source domain multitasking model within that client is built based on the loss function of each task branch network of the source domain multitasking model.

Because the source domain multitask evaluation model belongs to a three-task network, each task has a respective loss function, in order to realize joint training, the evaluation effect of the three-task network is optimal, and the loss function of the source domain multitask evaluation model can be constructed by combining the loss functions of the three-task network; when each client trains the source domain state evaluation model locally, the source domain is used for more than any Loss function Loss (x _k ,l _k ,s _k ,e _k ) The target is subjected to iterative updating by the minimization, and a specific calculation formula is as follows:

Loss(x _k ,l _k ,s _k ,e _k )＝Loss _RUL (x _k ,l _k )+λLoss _HS (x _k ,s _k )

+μLoss _RA (x _k ,e _k )；

the calculation formula of the loss function of the three branch networks is as follows:

/>

wherein lambda and mu are weight coefficients, and are determined by the contribution degree of different losses, loss _RA 、Loss _RUL And Loss of _HS Respectively evaluating the loss function of the predicted task branch network, the loss function of the residual life predicted task branch network and the loss function of the health stage division predicted task branch network for the reliability in the source domain multitask evaluation model; f (f) _RUL (*)、f _RA (*)、The method comprises the steps of respectively obtaining a residual life predicted value, a reliability evaluation predicted value and a health stage division predicted value, wherein C is the health stage class number of the equipment, in is a logarithmic function, and K is the number of samples constructed by different sample equipment In different operation periods; l (L) _k 、e _k 、s _k The remaining life label, the reliability evaluation label, and the health staging label for sample k, respectively.

Then, after the trained source domain multitasking evaluation model is obtained, an initial target domain coding model and an initial target domain multitasking model may be built based on the trained source domain multitasking evaluation model, and the specific steps include:

initializing an original target domain coding model based on model parameters of a source domain coding model of the trained source domain multitasking model to obtain an original target domain coding model, and initializing the original target domain multitasking model based on model parameters of the source domain multitasking model of the trained source domain multitasking model to obtain the original target domain multitasking model.

Then, performing federal learning training of the countermeasure field self-adaption on the initial target field coding model to obtain a target field coding model, wherein the specific steps are as follows:

optionally, after the source domain multitasking model is trained, in order to better transfer the features from the source domain to the target domain, the trained source domain multitasking evaluation model is subjected to an antagonistic adaptation training. In this process, two encoders are used, one of which is used to map source domain data and the other of which is used to map target domain data. And for each client in the federal learning framework, in the local training process, the self-adaptive training is carried out on the initial target domain coding model by keeping the weight of the source domain coding model unchanged and adopting the domain label under the source domain and the domain label under the target domain, so as to realize the alignment of the feature numbers of the source domain and the target domain.

The adaptive training process of the antagonism adaptation is to map an adaptive training data set under a source domain (namely, a data set containing a domain label and a time-frequency domain characteristic of sample equipment) with an adaptive training data set under a target domain through coding models under respective domains, and then use a discriminator to identify domain categories. By taking reference to the idea of a generative antagonism network (Generative Adversarial Nets, GANs), adaptive training is performed by minimizing the distance between the feature mapping distribution under the source domain and the feature mapping distribution under the target domain as a target.

In the process of training the initial target domain coding model locally by each client, the field label of the sample equipment and the coding characteristics of the sample equipment output by the target domain coding model and the coding characteristics of the sample equipment output by the pre-trained source coding model are subjected to field classification by the discriminator, the loss function constructed by the output field recognition result is the minimum as a target, and the initial target domain coding model is subjected to federal learning training.

Local adaptation domain training for initial target domain coding model within each client: inputting the time domain characteristics and the frequency domain characteristics of the first sample equipment under the source domain in the self-adaptive training data set in the client into the trained source coding model to obtain source domain coding characteristics; and inputting the time domain characteristics and the frequency domain characteristics of the second sample equipment in the target domain in the self-adaptive training data set in the client into the initial target domain coding model updated based on the global model parameters corresponding to the last model training to obtain the target domain coding characteristics. Then, the target domain coding feature and the source domain coding feature are input into a discriminator, the discriminator performs countermeasure learning on the target domain coding feature and the source domain coding feature to identify a domain type predicted value to which the sample equipment belongs, a loss function corresponding to the initial target domain coding model is obtained according to the domain type predicted value to which the sample equipment belongs and a domain label of the sample equipment, and further local training of countermeasure domain adaptation is performed on the initial target domain coding model according to the loss function to obtain a local training result.

Then, based on the target domain coding model, performing federal learning training of transfer learning on the initial target domain multitask model to obtain the target domain multitask model, wherein the specific steps comprise:

and performing fine tuning training of federal learning on the initial target domain multitask model based on the coding features obtained by inputting the time domain features and the frequency domain features of each second sample device in the target domain training set in the first client and the corresponding multitask labels, and the coding features obtained by inputting the time domain features and the frequency domain features of each second sample device in the target domain training set in each second client and the corresponding multitask labels, so as to obtain the target domain multitask model.

According to the method provided by the embodiment, the target domain multitask evaluation model is obtained through training by combining the federal learning and the countermeasure domain self-adaptive method, on one hand, the problem of inconsistent data distribution can be effectively solved through the countermeasure domain self-adaptive method, and the prediction performance of the target domain multitask evaluation model on the target domain can be improved through learning the sharing characteristic between the source domain and the target domain data, so that more accurate equipment health state evaluation is realized. On the other hand, through the federal learning mode, namely training data in a plurality of local models, uploading the local models to the cloud for integration, and finally generating a global model. In the process, each local model can only access own data, and only can access model parameters in the interaction process, so that the privacy of the data can be ensured, the problem of data privacy is solved, the generalization capability of the model can be effectively improved, and the accuracy of the equipment state evaluation result is further effectively improved.

In some embodiments, the loss function corresponding to the initial target domain coding model in the first client is jointly constructed based on the correctness probability value of the domain prediction result corresponding to each first sample device in the source domain training set in the first client and the correctness probability value of the domain prediction result corresponding to each second sample device in the target domain training set.

Optionally, the initial target encoding module within the first client is configured to train the target to minimize a similarity distance between the encoded features of the first sample device under the source domain and the encoded features of the second sample device under the target domain.

By means of the training method of the countermeasure adaptation, the source domain data X can be used with two encoders _s And target domain data X _t Mapping is performed and the two domains are classified using a discriminator. By learning in thought of generating countermeasure network by reference, source domain distribution M is realized _s (X _s ) Distribution M with target domain _t (X _t ) The distance between them is minimized. In particular, the encoder and discriminator are continually optimized by training to achieve a high level of performance in bothData transfer is performed between the domains.

In the countermeasure domain adaptive training process, the domain classification problem of the source domain and the target domain can be regarded as a classification problem, for example, the source domain label is set to 1, the target domain label is set to 0, and then the discriminator is utilized to distinguish whether the mapped data belongs to the source domain or the target domain. In this way, the domain identification problem between the source domain and the target domain in the cross-domain task is regarded as a binary classification task, so that the migration learning is effectively performed.

The discriminator is expected to be able to accurately distinguish the coded features mapped by the coding module, so the first optimization objective of the discriminator D is to minimize the following function:

wherein the method comprises the steps ofRepresenting that the discriminator predicts a correct probability value for the domain corresponding to the first sample device under the source domain; />Representing that the discriminator predicts a correct probability value for the domain corresponding to the second sample device under the source domain; />Is a sample under the source domain; />Is a sample under the target domain; />And->Respectively, the coding features of the time-frequency domain features of the first sample device under the source domain output by the trained source coding module toAnd the initial target coding module outputs the coding characteristics of the time-frequency domain characteristics of the second sample equipment under the target domain; d is the output result of the discriminator. />

Unlike the target of the discriminator, the target domain coding model aims to reduce the distance between data from different domains by mapping, which makes it difficult to distinguish the mapped data of the discriminator. Thus, the target domain coding model learns domain invariant features. In the initial stage of training, the data mapped by the target domain coding model may be poor in quality, so that the identifier can easily and correctly judge the distribution characteristics of the operating parameters of the second sample equipment from the target domain, thereby causing Saturation is reached and sufficient gradients cannot be provided for training of the target domain coding model. Thus, a second optimization objective of the discriminator D may be added to minimize the following function:

wherein L is _m Representing the domain loss function of the sample device under the target domain after domain adaptation.

Combining the two loss functions to obtain a loss function L corresponding to the initial target domain coding model _DA The following are provided:

min _D,Mt L _DA (X _s ,X _t ,M _s )＝L _disc (X _s ,X _t ,M _s ,M _t )+L _m (X _t ,D)。

in order to further verify the validity of the device state evaluation method of the present embodiment, the present embodiment uses different rolling bearing data under different working conditions in the PHM2012 (Prognostics Health Management,2012 fault prediction and health management) data set as a source domain and a target domain, and compares the source domain and the target domain with other models and existing methods to verify the validity of the method.

Table 1 error comparison table for reliability evaluation

Table 2 error comparison table for residual life prediction

As shown in tables 1 and 2, and fig. 6 and 7, there are error comparison tables and error charts for reliability evaluation and residual life prediction, respectively. By adopting the bearing 1-1 under the working condition 1 as sample equipment under the source domain, other bearings under the working condition 1 and other bearings under the working condition are adopted as sample equipment under the target domain, and a plurality of groups of migration tasks are designed; wherein, A-B represents the effect obtained by migrating on Bearing1_7 with Bearing1_1 as the source domain and Bearing1_2-Bearing1_6 as the target domain in PHM2012 data set, A-C represents the effect obtained by migrating on Bearing2_6 with Bearing1_1 as the source domain and Bearing2_1-Bearing2_5 as the target domain in PHM2012 data set, and A-D represents the effect obtained by migrating on Bearing3_3 with Bearing1_1 as the source domain and Bearing3_1-Bearing3_2 as the target domain in PHM2012 data set. And using RMSE (Root Mean Square Error ), MAE (Mean Absolute Error, mean absolute error) as evaluation index; as can be seen from the evaluation result, the reliability evaluation and the residual life prediction error of the method based on the adaptive migration learning training in the countermeasure field provided in this embodiment are lower than those of the method without migration training.

Wherein, the expressions of RMSE and MAE are:

where n is the number of sample devices, RUL _real For remaining life label, RUL _pred Is a residual life prediction value.

The errors of the two regression tasks of reliability evaluation and life prediction are obviously reduced after feature migration is not difficult to see from the evaluation index result and the error map, and the effectiveness of the method is proved.

TABLE 3 error contrast table for healthy phase partition prediction

Health staging	A-B	A-C	A-D
				Accuracy	0.922	0.947	0.950
Recall	0.901	0.938	0.958
				F1 Score	0.933	0.941	0.955

Tables 3 and 8 show the results of the migrated health-stage-divided evaluation index and the health-stage-divided simulation diagram, respectively. Where Accuracy, recall, and F1 Score are used, the expressions are:

wherein the method comprises the steps ofFor accuracy, TP, TN, FP, FN represents the number of samples of true positive, true negative, false positive and false negative, respectively.

Table 4 shows three evaluation index scores for the proposed method and conventional feature migration, such as TCNN (Transfer Convolutional Neural Network, convolutional neural network based on migration learning), TBiLSTM (Transfer Bi-Directional Long Short-Term Memory network based on migration learning), and TCA (Transfer Component Analysis, migration component analysis), where the Score expression is:

Wherein RUL _real For remaining life label, RUL _pred A predicted value for remaining life; a is that _i Score coefficient for sample device i; e (E) _i Is the error coefficient of the sample device i.

Table 4 comparison of evaluation index scores of various migration methods

It can be seen from table 4 that the effect of the method proposed by the present embodiment is due to the conventional method.

Table 5 comparison table of evaluation index scores for reliability evaluation under different transfer learning algorithms

As shown in table 5 and table 6, two regression tasks for reliability evaluation and residual life prediction are shown, and the effects of the federal learning provided in this embodiment are improved to a certain extent compared with the effects of the other two methods by performing post-training effect comparison on the basis of the Centralized learning (Centralized learning) and the federal average algorithm (Fedavg) provided in the prior art.

Table 6 evaluation index score comparison table for residual life prediction under different migration learning algorithms

The following describes the system for evaluating the health state of a device for multi-task learning and privacy protection provided by the present invention, and the system for evaluating the health state of a device for multi-task learning and privacy protection described below and the method for evaluating the health state of a device for multi-task learning and privacy protection described above can be referred to correspondingly with each other.

Fig. 9 is a schematic structural diagram of a device health status assessment system for multi-task learning and privacy protection provided in this embodiment; as shown in fig. 9, the system is applied to any first client of a plurality of clients, including:

the first extraction module 901 is configured to perform singular value decomposition on a historical operation parameter of each sample device in the first client to obtain a time domain feature and a frequency domain feature corresponding to the historical operation parameter; the sample devices comprise a first sample device under a source domain and a second sample device under a target domain;

the second extraction module 902 is configured to perform feature reconstruction on the time domain features and the frequency domain features to obtain a health indicator of each sample device;

the marking module 903 is configured to perform mutation detection on the health indicator, and obtain a multi-task label of each sample device according to a mutation detection result; the multi-task labels comprise a residual life label, a reliability evaluation label and a health stage division label;

the building module 904 is configured to build a source domain training set based on the multi-task tag, the time domain feature, and the frequency domain feature of the first sample device in the source domain, and build a target domain training set based on the multi-task tag, the time domain feature, and the frequency domain feature of the second sample device in the target domain;

The training module 905 is configured to perform adaptive federal learning training in the challenge domain based on the source domain training set and the target domain training set in the first client, and the source domain training set and the target domain training set in each of the second clients except the first client, to obtain a target domain multitask evaluation model;

According to the equipment health state evaluation system provided by the embodiment, on one hand, the training data set is constructed efficiently and accurately through singular value decomposition and mutation inspection, and the accuracy and the efficiency of the training of a subsequent target domain multitask evaluation model are improved; on the other hand, the field self-adaptive training is carried out through the training data set to realize the feature mapping and migration of the source domain and the target domain, and the federal learning framework with stronger robustness is introduced to carry out the collaborative training of the target domain state evaluation model, so that the alignment of the source domain data and the target domain data is realized, the data privacy is protected, the expansion of the data set is realized, and the accuracy, the generalization and the robustness of the equipment health state evaluation are improved.

Fig. 10 illustrates a physical structure diagram of an electronic device, as shown in fig. 10, which may include: a processor 1001, a communication interface (Communications Interface) 1002, a memory 1003, and a communication bus 1004, wherein the processor 1001, the communication interface 1002, and the memory 1003 perform communication with each other through the communication bus 1004. The processor 1001 may call logic instructions in the memory 1003 to perform the device health assessment method of multi-task learning and privacy protection provided by the above embodiments.

Further, the logic instructions in the memory 1003 described above may be implemented in the form of software functional units and sold or used as a separate product, and may be stored in a computer readable storage medium. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

In another aspect, the present invention also provides a computer program product comprising a computer program storable on a non-transitory computer readable storage medium, the computer program, when executed by a processor, being capable of performing the method of assessing device health status for multi-tasking learning and privacy protection provided by the methods described above.

In yet another aspect, the present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, is implemented to perform the method of assessing device health status for multitasking learning and privacy protection provided by the methods described above.

The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A device health status assessment method for multitasking learning and privacy protection, applied to a first client of a plurality of clients, comprising:

2. The method for evaluating the health state of a device for multi-task learning and privacy protection according to claim 1, wherein the performing singular value decomposition on the historical operating parameters of each sample device in the first client to obtain the time domain feature and the frequency domain feature corresponding to the historical operating parameters comprises:

for each of the sample devices, performing the following operations:

3. The method for evaluating the health status of a device for multi-task learning and privacy protection according to claim 1, wherein the performing feature reconstruction on the time domain features and the frequency domain features to obtain the health index of each sample device comprises:

4. The method for evaluating the health status of a device for multi-task learning and privacy protection according to claim 1, wherein said performing a mutation test on the health index and obtaining a multi-task tag of each sample device according to a mutation test result comprises:

5. The method for assessing the health of a device for multi-tasking learning and privacy protection of any of claims 1-4 wherein the target domain multi-tasking assessment model comprises a target domain coding model and a target domain multi-tasking network;

6. The method for evaluating the health status of a device for multi-task learning and privacy protection according to claim 5, wherein the performing the federal learning training for challenge domain adaptation based on the source domain training set and the target domain training set in the first client and the source domain training set and the target domain training set in each of the second clients other than the first client to obtain the target domain multi-task evaluation model comprises:

7. The method for evaluating the health status of a device for multi-task learning and privacy protection according to claim 6, wherein the loss function corresponding to the initial target domain coding model in the first client is jointly constructed based on the correctness probability value of the domain prediction result corresponding to each first sample device in the source domain training set in the first client and the correctness probability value of the domain prediction result corresponding to each second sample device in the target domain training set.

8. A device health status assessment system for multitasking learning and privacy protection, applied to any first client of a plurality of clients, comprising:

9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of device health assessment for multi-task learning and privacy protection as claimed in any one of claims 1 to 7 when the program is executed by the processor.

10. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the method of device health state assessment for multi-task learning and privacy protection as claimed in any one of claims 1 to 7.