Bearing diagnostic method, system, medium and device for multi-source subdomain negative migration inhibition

Abstract

The present invention relates to the technical field of bearing fault diagnosis, and discloses a bearing diagnosis method, system, medium and equipment with multi-source subdomain negative migration suppression, including: obtaining source domain data and target domain data corresponding to the bearing, and bearing data to be tested, constructing a fault diagnosis model including a feature extractor, a domain discriminator, a specific domain classifier including multiple classifiers, and a local maximum mean difference module, the domain discriminator aligns the global distribution of all domain features through domain adversarial alignment and then inputs the specific domain classifier, the specific domain classifier outputs the predicted label of the source domain and the pseudo label of the target domain, the local maximum mean difference module combines the pseudo labels output by all classifiers and the local maximum mean difference to align the subdomains of all domains; the bearing data to be tested is input into the trained fault diagnosis model to obtain the fault diagnosis result of the bearing. The present invention can solve the problem of negative migration in the source domain and improve the accuracy of bearing fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing diagnosis method, system, medium and equipment for multi-source sub-domain negative migration suppression. Background Art

[0002] Rolling bearings play an important role in modern industry and are favored by customers for their high-precision and high-speed working characteristics. After a long period of operation, the inner ring, outer ring and roller of a rolling bearing may fail. Timely identification and diagnosis of these failures in rolling bearings can help to immediately find the problem rolling bearing and reduce the economic losses caused by bearing failures.

[0003] At present, when aligning data, the intelligent fault diagnosis model based on rolling bearings usually uses two methods: aligning the global distribution of data and aligning the subdomains of a single source domain and target domain. The process of aligning the global distribution of data, that is, global domain adaptation, is as follows: Figure 1 As shown in , this approach will cause confusion in the data near the class boundary, so there is a way to align the subdomains of a single source domain and a target domain. The process of aligning the subdomains of a single source domain and a target domain, that is, adjusting the subdomain adaptation of the same category in the two domains, is as follows Figure 2 As shown in the figure, this method can improve the fault diagnosis performance. However, it is difficult to unify the subdomain invariant features in the feature spaces of multiple source domains and target domains. There are large distribution differences in the multiple source domain data of rolling bearings under different working conditions, so it is difficult to extract the key features of each source domain. At the same time, this method only aligns the subdomains of a single source domain and the target domain without considering the problem of subdomain distribution alignment between source domains, making it difficult for the source domain with a large distribution difference from the target domain to align with the target domain and produce negative migration, resulting in large differences in the classification accuracy of different classifiers for the same target domain data, affecting the accuracy of fault identification. Summary of the invention

[0004] To this end, the technical problem to be solved by the present invention is to overcome the deficiencies in the prior art and provide a bearing diagnosis method, system, medium and equipment for suppressing negative migration in multiple source subdomains, which can solve the problem of negative migration in the source domain and improve the accuracy of bearing fault diagnosis.

[0005] In order to solve the above technical problems, the present invention provides a bearing diagnosis method for multi-source sub-domain negative migration suppression, comprising:

[0006] Acquire source domain data and target domain data corresponding to the bearing and perform preprocessing, and construct a fault diagnosis model including a feature extractor, a domain discriminator, a specific domain classifier, and a local maximum mean difference module, wherein the specific domain classifier includes multiple classifiers;

[0007] The feature extractor extracts the features of the preprocessed source domain data and target domain data and inputs them into the domain discriminator. The domain discriminator aligns the global distribution of all domain features through domain adversarial analysis and then inputs them into the specific domain classifier. The specific domain classifier outputs the predicted label of the source domain and the pseudo label of the target domain. The local maximum mean difference module combines the pseudo labels output by all classifiers and the local maximum mean difference to align the subdomains of all domains.

[0008] The preprocessed source domain data and target domain data are input into the fault diagnosis model to train the fault diagnosis model to obtain a trained fault diagnosis model, the bearing data to be tested is obtained for preprocessing, and the preprocessed bearing data to be tested is input into the trained fault diagnosis model to obtain the fault diagnosis result of the bearing.

[0009] In one embodiment of the present invention, the local maximum mean difference module combines the pseudo labels output by all classifiers and the local maximum mean difference to align the subdomains of all domains, specifically:

[0010] The source domain features obtained after the preprocessed source domain data passes through the specific domain classifier are denoted as Z s , the target domain feature obtained after the preprocessed target domain data passes through the specific domain classifier is Z t ; The true label of the source domain is y s , is the true label of the i-th source domain data, N1 is the number of source domain data; Z t The pseudo label of the target domain obtained after the softmax function operation is is the pseudo label of the jth target domain, and N2 is the number of target domain data;

[0011] The source domain feature Z s , target domain feature Z t , the true label of each source domain data Pseudo labels for each target domain Input the local maximum mean difference module for subdomain alignment, and calculate the mean square error between the pseudo labels of the target domain output by all classifiers. Close the distribution of each pseudo-label and finally align the subdomain distributions of different source domains.

[0012] In one embodiment of the present invention, when the fault diagnosis model is trained, according to the source domain feature Z s and the target domain feature Z t Compute domain classification loss The local maximum average difference classification loss between source domain features and target domain features is calculated based on the pseudo labels of the target domain output by all classifiers Predicted labels based on the source domain and the true label y s Calculating the classifier loss

[0013] Combined with the mean square error Domain Classification Loss Local maximum mean discrepancy classification loss Classifier Loss A total loss function is constructed, and the training is terminated until the total loss function converges or the maximum number of training times is reached, thereby obtaining a trained fault diagnosis model.

[0014] In one embodiment of the present invention, when the fault diagnosis model is trained, the mean square error is back-propagated Domain Classification Loss Local maximum mean discrepancy classification loss Classifier Loss The model parameters of the fault diagnosis model are updated, and the predicted labels of the source domain and the pseudo labels of the target domain are obtained by forward propagation.

[0015] In one embodiment of the present invention, the mean square error The calculation method is:

[0016]

[0017] Where N represents the number of classifiers, G c,i () represents the i-th domain-specific classifier, G lf,i () represents the i-th local feature extractor, G gf () represents the global feature extractor, G c,j () represents the jth domain-specific classifier, G lf,j () represents the jth local feature extractor, G gf () represents the global feature extractor, x t The target domain features representing the input;

[0018] The domain classification loss The calculation method is:

[0019]

[0020] Where M represents the number of data in each training batch, d i represents the real domain label of the input data, σ is the sigmoid function, represents the i-th source domain input feature of the domain discriminator, represents the i-th target domain input feature of the domain discriminator; G d () represents the domain discriminator, G lf () represents the local feature extractor, G gf () represents the global feature extractor.

[0021] In one embodiment of the present invention, the local maximum average difference classification loss The calculation method is:

[0022]

[0023] Where C represents the total number of fault categories, n s Represents the number of input source domain samples, represents the weight of the i-th source sample belonging to the c-th class, represents the weight of the jth source sample belonging to the cth class, k() represents the kernel function, represents the output feature of the i-th source domain sample in the l-th layer of the network, represents the output feature of the jth source domain sample in the lth layer of the network; n t represents the number of input target domain samples, represents the weight of the i-th target sample belonging to the c-th class, represents the weight of the jth target sample belonging to the cth class, represents the output feature of the i-th target domain sample in the l-th layer of the network, Represents the output feature of the jth target domain sample in the lth layer of the network;

[0024] The classifier loss The calculation method is:

[0025]

[0026] Where M represents the number of data in each training batch. represents the true label of the i-th source domain data, is the predicted label of the i-th source domain data, C represents the total number of fault categories, and c represents one of the categories. represents the label vector whose true label is c; σ is the softmax function, G c ( ) represents a specific domain classifier, G lf () represents the local feature extractor, G gf () represents the global feature extractor.

[0027] In one embodiment of the present invention, the feature extractor includes a global feature extractor and a local feature extractor, the global feature extractor has residual and channel expansion downsampling functions, the local feature extractor includes depthwise separable convolution and convolutional attention, and the convolutional attention includes a channel attention mechanism and a spatial attention mechanism.

[0028] The present invention also provides a bearing diagnosis system for multi-source sub-domain negative migration inhibition, comprising:

[0029] A data acquisition module is used to acquire source domain data and target domain data corresponding to the bearing and perform preprocessing, and acquire the bearing data to be tested for preprocessing;

[0030] A fault diagnosis model construction module is used to construct a fault diagnosis model including a feature extractor, a domain discriminator, a specific domain classifier, and a local maximum mean difference module, wherein the specific domain classifier includes a plurality of classifiers; the feature extractor extracts features of the preprocessed source domain data and target domain data and inputs the features into the domain discriminator, the domain discriminator aligns the global distribution of all domain features through domain adversarial analysis and then inputs the features into the specific domain classifier, the specific domain classifier outputs a predicted label of the source domain and a pseudo label of the target domain, and the local maximum mean difference module combines the pseudo labels output by all classifiers and the local maximum mean difference to align subdomains of all domains;

[0031] A training module, used for inputting the preprocessed source domain data and target domain data into the fault diagnosis model to train the fault diagnosis model to obtain a trained fault diagnosis model;

[0032] The fault diagnosis module is used to input the preprocessed bearing data to be tested into the trained fault diagnosis model to obtain the bearing fault diagnosis result.

[0033] The present invention also provides a computer-readable storage medium, and when the computer program is executed by a processor, the bearing diagnosis method for multi-source sub-domain negative migration inhibition as described above is implemented.

[0034] The present invention also provides a bearing diagnostic device with multi-source sub-domain negative migration inhibition, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the bearing diagnostic method with multi-source sub-domain negative migration inhibition when executing the computer program.

[0035] The above technical solution of the present invention has the following advantages compared with the prior art:

[0036] The present invention sets a specific domain classifier and a local maximum mean difference module including multiple classifiers, combines the pseudo labels output by all classifiers and the local maximum mean difference to align the subdomains of all domains, and uses the decision boundary of the specific domain to promote the output alignment of different classifiers, thereby aligning the subdomains between various source domains and between the source domain and the target domain to solve the source domain negative migration problem, so as to better extract the key features of each source domain and improve the accuracy of bearing fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0038] Figure 1 It is a schematic diagram of aligning the global distribution of data.

[0039] Figure 2 It is a schematic diagram of aligning subdomains of a single source domain and a target domain.

[0040] Figure 3 It is a schematic diagram of multi-source sub-domain adaptation in the present invention.

[0041] Figure 4 is a flow chart of the method of the present invention.

[0042] Figure 5 It is a schematic diagram of aligning all domain subdomains by combining the local maximum mean difference guided by multiple pseudo labels in the present invention.

[0043] Figure 6 It is a structural schematic diagram of the present invention.

[0044] Figure 7 It is a step diagram of the method of the present invention.

[0045] Figure 8 It is a diagram showing the effect of using the pseudo labels of the target domain output by all classifiers to guide LMMD training in an embodiment of the present invention.

[0046] Fig. 9 This is a diagram showing the effect of using only the pseudo-label of the target domain output by a specific domain classifier corresponding to the source domain to guide model training in an embodiment of the present invention.

[0047] Fig.10 is a feature visualization result diagram of the T2 task in an embodiment of the present invention.

[0048] Fig.11 is a feature visualization result diagram of the T1000 task in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0050] Embodiment 1

[0051] Reference Figure 1-Figure 7 As shown, the present invention discloses a bearing diagnosis method for multi-source subdomain negative migration inhibition, comprising the following steps:

[0052] S1: Obtain the source domain data and target domain data corresponding to the bearing and perform preprocessing to build a fault diagnosis model (denoted as MSA-MSDA).

[0053] S1-1: Obtain source domain data and target domain data corresponding to the bearing for preprocessing, the preprocessing includes: performing fast Fourier transform (FFT), normalization and dimension increase operations on the sample data. In this embodiment, each original data sample is processed into three-dimensional data that is convenient for convolution operation.

[0054] S1-2: Construct a fault diagnosis model (denoted as MSA-MSDA) including a global feature extractor (G), a local feature extractor (L), a domain discriminator (D), a specific domain classifier (C), and a local maximum mean difference module, wherein the specific domain classifier includes multiple classifiers.

[0055] In this embodiment, the specific structure of the fault diagnosis model is shown in Table 1. Taking the number of samples input into the fault diagnosis model in each batch as 64 as an example, the sample dimensions after preprocessing are [64, 1, 32, 32]. Among them, the global feature extractor has residual and channel expansion downsampling functions, which can better capture the rich global feature information in the data and make the sampled features more representative, thereby improving the understanding and generalization ability of the model; the local feature extractor includes depthwise separable convolution and convolutional attention, and the convolutional attention includes a channel attention mechanism and a spatial attention mechanism, thereby increasing the model's ability to extract local key features.

[0056] Table 1 Specific structure of fault diagnosis model

[0057]

[0058]

[0059] The global feature extractor projects the data into a common feature space and then inputs the data into the local feature extractor. The local feature extractor extracts the key information of a specific domain and then inputs the data into the domain discriminator. The domain discriminator aligns the global distribution of all domain features through domain adversarial analysis and then inputs the data into the specific domain classifier. The specific domain classifier outputs the predicted label of the source domain and the pseudo label of the target domain. The local maximum mean difference module combines the pseudo label and the local maximum mean difference (LMMD) to align the subdomains of all domains.

[0060] Figure 5 N in the above formula represents the number of classifiers. When N=1, it is the method in the prior art that only the pseudo-label output by a single classifier is used to guide LMMD. Only the subdomains of a single source domain and a target domain are aligned without considering the problem of subdomain distribution alignment between source domains, which will produce negative migration. In view of this, the present invention sets a specific domain classifier including multiple classifiers. The source domain feature obtained after the preprocessed source domain data passes through the specific domain classifier is denoted as Z s , the target domain feature obtained after the preprocessed target domain data passes through the specific domain classifier is Z t ; The true label of the source domain is y s , is the true label of the i-th source domain data; N1 is the number of source domain data for each batch of input model training; Z t The pseudo label of the target domain obtained after the softmax function operation is is the pseudo label of the jth target domain, and N2 is the number of target domain data for each batch of input model training. s , target domain feature Z t , the true label of each source domain data Pseudo labels for each target domain The local maximum mean difference module is input for subdomain alignment (SDA), and the mean square error between the pseudo labels of the target domain output by all classifiers is calculated. The distribution of each pseudo-label is brought closer, and finally the sub-domain distribution of different source domains is aligned to promote the correct classification of each classifier. The local maximum mean difference module in this embodiment combines the influence of sample categories on the kernel function weights, and utilizes fine-grained information in domain adaptation, which can better align the sub-domain distribution of data; by training all classifier pseudo-labels, the pseudo-labels of the target domain with higher classification accuracy can guide the learning of lower classifiers; by bringing the sub-domain distribution between target domains closer, the sub-domain distribution of different source domains is aligned to avoid negative transfer of the source domain.

[0061] S2: Input the preprocessed source domain data and target domain data into the fault diagnosis model to train the fault diagnosis model to obtain a trained fault diagnosis model. The preprocessed source domain data is represented by D sj , D sj ={X sj ,Y sj}, where X sj is the feature of the source domain data, Y sj is the label of the source domain data, and the feature representation of the preprocessed target domain data is X t , nt is the number of target domain data.

[0062] When training the fault diagnosis model, the forward propagation obtains the predicted label of the source domain and the pseudo-labels of the target domain Calculate the source domain feature Z s and the target domain feature Z t Domain classification loss Utilize the pseudo labels of the target domain output by all classifiers Calculate the local maximum average difference classification loss between source domain features and target domain features Compute the predicted labels for the source domain and the true label y s The classifier loss Compute the pseudo labels of the target domain output by all classifiers The mean square error between the two Combined with the mean square error Domain Classification Loss Local maximum mean discrepancy classification loss Classifier Loss Construct total loss function and back propagate The model parameters of the fault diagnosis model are updated until the total loss function converges or the maximum number of training times is reached, and the training is terminated to obtain a trained fault diagnosis model.

[0063] Specifically, the mean square error The calculation method is:

[0064]

[0065] Among them, G c,i () represents the i-th domain-specific classifier, G lf,i () represents the i-th local feature extractor, G gf () represents the global feature extractor, G c,j () represents the jth domain-specific classifier, G lf,j () represents the jth local feature extractor, G gf() represents the global feature extractor, x t Represents the target domain features of the input.

[0066] Domain Classification Loss The calculation method is:

[0067]

[0068] Where M represents the number of data in each training batch, d i represents the real domain label of the input data, σ is the sigmoid function, represents the i-th source domain input feature of the domain discriminator, G represents the i-th target domain input feature of the domain discriminator. d () represents the domain discriminator, G lf () represents the local feature extractor, G gf () represents the global feature extractor.

[0069] Local maximum mean discrepancy classification loss The calculation method is:

[0070]

[0071] Where C represents the total number of fault categories, n s Represents the number of input source domain samples, represents the weight of the i-th source sample belonging to the c-th class, represents the weight of the jth source sample belonging to the cth class, k() represents the kernel function, represents the output feature of the i-th source domain sample in the l-th layer of the network, represents the output feature of the jth source domain sample in the lth layer of the network; n t represents the number of input target domain samples, represents the weight of the i-th target sample belonging to the c-th class, represents the weight of the jth target sample belonging to the cth class, represents the output feature of the i-th target domain sample in the l-th layer of the network, Represents the output features of the j-th target domain sample in the l-th layer of the network.

[0072] and The calculation method is as follows:

[0073]

[0074]

[0075] Where N1 is the number of different source domains, represents the source domain dataset, represents the target domain dataset; n s Represents the number of input source domain samples, n t Represents the number of target domain samples of the input; Represents a dataset from the source domain The jth data feature of Represents a dataset from the source domain The jth label of Represents a dataset from the target domain The jth data feature of Represents a dataset from the target domain The jth pseudo label of Represents the input source domain sample The corresponding true label of the source domain data in the kth field The cth element of Represents the input source domain sample The corresponding true label of the source domain data in the kth field the cth element of ; Represents the input target domain sample The corresponding pseudo label output by the k-th domain-specific classifier Item c of Represents the input target domain sample The corresponding pseudo label output by the k-th domain-specific classifier Item c of .

[0076] Classifier Loss The calculation method is:

[0077]

[0078] Where M represents the number of data in each training batch. represents the true label of the i-th source domain data, is the predicted label of the i-th source domain data, C represents the total number of fault categories, and c represents one of the categories; The label vector represents the real label c. In this embodiment, C = y s ={0, 1, 2, ... 9}, c∈C, that is, c = 0, 1, 2, ..., 9; σ is the softmax function, G c () represents a specific domain classifier, G lf () represents the local feature extractor, G gf () represents the global feature extractor.

[0079] The calculation method of the total loss function L is:

[0080]

[0081] S3: obtaining the bearing data to be tested for preprocessing, and inputting the preprocessed bearing data to be tested into the trained fault diagnosis model to obtain the fault diagnosis result of the bearing.

[0082] Embodiment 2

[0083] The present invention also discloses a bearing diagnosis system with multi-source sub-domain negative migration suppression, comprising a data acquisition module, a fault diagnosis model construction module, a training module and a fault diagnosis module.

[0084] A data acquisition module is used to acquire source domain data and target domain data corresponding to the bearing and perform preprocessing, and acquire the bearing data to be tested for preprocessing;

[0085] A fault diagnosis model construction module is used to construct a fault diagnosis model including a feature extractor, a domain discriminator, a specific domain classifier, and a local maximum mean difference module, wherein the specific domain classifier includes a plurality of classifiers; the feature extractor extracts features of the preprocessed source domain data and target domain data and inputs the features into the domain discriminator, the domain discriminator aligns the global distribution of all domain features through domain adversarial analysis and then inputs the features into the specific domain classifier, the specific domain classifier outputs a predicted label of the source domain and a pseudo label of the target domain, and the local maximum mean difference module combines the pseudo labels output by all classifiers and the local maximum mean difference to align subdomains of all domains;

[0086] A training module, used for inputting the preprocessed source domain data and target domain data into the fault diagnosis model to train the fault diagnosis model to obtain a trained fault diagnosis model;

[0087] The fault diagnosis module is used to input the preprocessed bearing data to be tested into the trained fault diagnosis model to obtain the bearing fault diagnosis result.

[0088] Embodiment 3

[0089] The present invention also discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the bearing diagnosis method for suppressing negative migration of multiple source subdomains in embodiment one is implemented.

[0090] Embodiment 4

[0091] The present invention also discloses a device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the bearing diagnosis method for suppressing negative migration of multiple source subdomains in embodiment one is implemented.

[0092] Compared with the prior art, the present invention has the following beneficial effects:

[0093] 1. The present invention sets a specific domain classifier including multiple classifiers and a local maximum mean difference module, combines the pseudo labels output by all classifiers and the local maximum mean difference to align the subdomains of all domains, performs subdomain alignment on the target domain features output by multiple specific domain classifiers, and then aligns the subdomain distributions of multiple source domains, thereby suppressing the negative migration of the source domain, so as to better extract the key features of each source domain and improve the accuracy of bearing fault diagnosis.

[0094] 2. The present invention uses a common domain discriminator to align the global distribution of key features of all domains, and then uses multiple pseudo-labels output by the classifier to guide the model to perform sub-domain alignment, thereby avoiding misleading the model by the pseudo-labels output by the classifier corresponding to the source domain with a large difference in data distribution with the target domain.

[0095] 3. The present invention adopts a multi-stage alignment strategy. First, the key features are globally aligned, then the source domain and the target domain are sub-domain aligned, and finally the mean square error between the pseudo labels output by all classifiers for the target domain is calculated, and the pseudo labels output by different classifiers are aligned to promote the correct classification of each classifier, thereby further improving the performance of rolling bearing fault diagnosis.

[0096] 4. The present invention increases the model's ability to extract key features by setting a global feature extractor and a local feature extractor, promotes global domain alignment, and reduces the model volume.

[0097] 5. The global feature extractor in the present invention has residual and channel expansion downsampling functions, which can better capture the rich global feature information in the data and make the sampled features more representative, thereby improving the understanding and generalization ability of the model.

[0098] 6. The local feature extractor in the present invention includes deep separable convolution and convolutional attention, which can increase the model's ability to extract local key features.

[0099] In order to further illustrate the beneficial effects of the present invention, in this embodiment, the method of the present invention and the prior art are used to perform a simulation experiment of bearing fault diagnosis.

[0100] First, the pseudo labels of the target domain output by all classifiers are used to guide LMMD training, and the pseudo labels of the target domain output by only a single domain-specific classifier corresponding to the source domain are used to guide model training. The effect of using the pseudo labels of the target domain output by all classifiers to guide LMMD training is shown in Figure 2. Figure 8 As shown, only the pseudo-labels of the target domain output by a single domain-specific classifier corresponding to the source domain are used to guide model training. Fig. 9 As shown. Figure 8-Figure 9It can be seen that each classifier of the model trained with multiple pseudo labels has a high classification accuracy for the target domain, while the accuracy of the classifier of the model trained with only a single pseudo label varies, and the final average accuracy is much lower than that of the case of multi-label guidance training.

[0101] Next, the fault diagnosis performance of the method of the present invention and the existing method are compared on two different data sets. The data sets used are the Soochow University data set (SCU) and the Shandong University of Science and Technology data set (SUST). The fault types and fault labels of the two data sets of SCU and SUST are shown in Table 2, where NOR represents normal data, IR1, IR2, and IR3 correspond to inner ring fault data with fault diameters of 0.2 mm, 0.4 mm, and 0.6 mm, respectively, BF1, BF2, and BF3 correspond to ball faults with fault diameters of 0.2 mm, 0.4 mm, and 0.6 mm, respectively, and OR1, OR2, and OR3 correspond to outer ring faults with fault diameters of 0.2 mm, 0.4 mm, and 0.6 mm, respectively.

[0102] Table 2 Fault types and fault labels of the two datasets SCU and SUST

[0103]

[0104] The migration tasks corresponding to SCU are shown in Table 3, and the migration tasks corresponding to SUST are shown in Table 4. T3 represents that the model learns 0, 1, and 2KN load data from the Suzhou University bearing data set to diagnose faults in 3KN load data, and the meanings of other rows of data are the same.

[0105] Table 3 SCU migration task table

[0106]

[0107] Table 4 SUST migration task list

[0108]

[0109] The rolling bearing vibration signal is preprocessed and input into the MSA-MSDA and other models of the present invention. The diagnostic results of different models on SCU are shown in Table 5, and the diagnostic results of different models on SUST are shown in Table 6. Among them, method 1 represents global alignment using maximum mean difference, method 2 represents global alignment using correlation alignment, method 3 represents not using global alignment method, method 4 represents not using domain adaptation method, method 5 represents only using class boundary alignment strategy, method 6 represents existing multi-source domain adaptation method M3SDA, method 7 represents existing multi-source domain adaptation method FTD-MSDA, and method 8 represents single-source domain adaptation method KMADA.

[0110] Table 5 Diagnosis results of different models on SCU (%)

[0111]

[0112] Table 6 Diagnosis results of different models on SUST (%)

[0113]

[0114]

[0115] It can be seen from Tables 5 and 6 that the MSA-MSDA model proposed in the present invention has the highest accuracy.

[0116] Finally, in this example, the clustering effect of MSA-MSDA in the present invention is analyzed, and two experiments, T2 in Table 3 and T1000 in Table 4, are selected for feature visualization. The feature visualization results of the T2 task are shown in FIG. Fig.10 As shown, the feature visualization results of the T1000 task are as follows Fig.11 As shown. The output layer of the local feature extractor is defined as layer 1, and the visualization result of the output feature corresponds to Fig.10 and Fig.11 (a) in the figure; the first convolution layer of the domain-specific classifier is defined as the second layer, and the visualization of the output features corresponds to Fig.10 and Fig.11 (b) in Figure 2; the second layer of partial convolution of the domain-specific classifier is defined as the third layer, and the visualization of the output features corresponds to Fig.10 and Fig.11 (c) in the figure; the third classification convolution of the domain-specific classifier is defined as the fourth layer, that is, the layer that uses multiple pseudo-labels to guide LMMD, and the visualization results of the output features correspond to Fig.10 and Fig.11 (d) in. Fig.10 and Fig.11 It can be seen that the data of the same fault category in the source domain and the target domain have not been clustered, the samples of each category in the two domains are well clustered, and the data of the same category in the source domain and the target domain are also clustered. The present invention can effectively achieve clustering.

[0117] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0118] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0119] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0120] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0121] Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from these are still within the protection scope of the invention.

Claims

1. A bearing diagnosis method with multi-source subdomain negative migration suppression, characterized in that: include: Acquire source domain data and target domain data corresponding to the bearing and perform preprocessing, and construct a fault diagnosis model including a feature extractor, a domain discriminator, a specific domain classifier, and a local maximum mean difference module, wherein the specific domain classifier includes multiple classifiers; The feature extractor extracts the features of the preprocessed source domain data and target domain data and inputs them into the domain discriminator. The domain discriminator aligns the global distribution of all domain features through domain adversarial analysis and then inputs them into the specific domain classifier. The specific domain classifier outputs the predicted label of the source domain and the pseudo label of the target domain. The local maximum mean difference module combines the pseudo labels output by all classifiers and the local maximum mean difference to align the subdomains of all domains. Input the preprocessed source domain data and target domain data into the fault diagnosis model to train the fault diagnosis model to obtain a trained fault diagnosis model, obtain the bearing data to be tested for preprocessing, input the preprocessed bearing data to be tested into the trained fault diagnosis model to obtain a bearing fault diagnosis result; The local maximum mean difference module combines the pseudo labels output by all classifiers and the local maximum mean difference to align the subdomains of all domains, specifically: The source domain features obtained after the preprocessed source domain data passes through the specific domain classifier are denoted as Z s , the target domain feature obtained after the preprocessed target domain data passes through the specific domain classifier is Z t ; The true label of the source domain is y s , is the true label of the i-th source domain data, N1 is the number of source domain data; Z t The pseudo label of the target domain obtained after the softmax function operation is is the pseudo label of the jth target domain, and N2 is the number of target domain data; The source domain feature Z s , target domain feature Z t , the true label of each source domain data Pseudo labels for each target domain Input the local maximum mean difference module for subdomain alignment, and calculate the mean square error between the pseudo labels of the target domain output by all classifiers. Bring the distribution of each pseudo-label closer and align the subdomain distributions of different source domains.

2. The bearing diagnosis method with multi-source sub-domain negative migration suppression according to claim 1 is characterized in that: When training the fault diagnosis model, according to the source domain feature Z s and the target domain feature Z t Compute domain classification loss The local maximum average difference classification loss between source domain features and target domain features is calculated based on the pseudo labels of the target domain output by all classifiers Predicted labels based on the source domain and the true label y s Calculating the classifier loss Combined with the mean square error Domain Classification Loss Local maximum mean discrepancy classification loss Classifier Loss A total loss function is constructed, and the training is terminated until the total loss function converges or the maximum number of training times is reached, thereby obtaining a trained fault diagnosis model.

3. The bearing diagnosis method with multi-source sub-domain negative migration suppression according to claim 2 is characterized in that: When training the fault diagnosis model, back propagating the mean square error Domain Classification Loss Local maximum mean discrepancy classification loss Classifier Loss The model parameters of the fault diagnosis model are updated, and the predicted labels of the source domain and the pseudo labels of the target domain are obtained by forward propagation.

4. The bearing diagnosis method with multi-source sub-domain negative migration suppression according to claim 2 is characterized in that: The mean square error The calculation method is: Where N represents the number of classifiers, G c,i () represents the i-th domain-specific classifier, G lf,i () represents the i-th local feature extractor, G gf () represents the global feature extractor, G c,j () represents the jth domain-specific classifier, G lf,j () represents the jth local feature extractor, G gf () represents the global feature extractor, x t The target domain features representing the input; The domain classification loss The calculation method is: Where M represents the number of data in each training batch, d i represents the real domain label of the input data, σ is the sigmoid function, represents the i-th source domain input feature of the domain discriminator, represents the i-th target domain input feature of the domain discriminator; G d () represents the domain discriminator, G lf () represents the local feature extractor, G gf () represents the global feature extractor.

5. The bearing diagnosis method with multi-source sub-domain negative migration suppression according to claim 2, characterized in that: The local maximum mean difference classification loss The calculation method is: Where C represents the total number of fault categories, n s Represents the number of input source domain samples, represents the weight of the i-th source sample belonging to the c-th class, represents the weight of the jth source sample belonging to the cth class, k() represents the kernel function, represents the output feature of the i-th source domain sample in the l-th layer of the network, represents the output feature of the jth source domain sample in the lth layer of the network; n t represents the number of input target domain samples, represents the weight of the i-th target sample belonging to the c-th class, represents the weight of the jth target sample belonging to the cth class, represents the output feature of the i-th target domain sample in the l-th layer of the network, Represents the output feature of the jth target domain sample in the lth layer of the network; The classifier loss The calculation method is: Where M represents the number of data in each training batch. represents the true label of the i-th source domain data, is the predicted label of the i-th source domain data, C represents the total number of fault categories, and c represents one of the categories. represents the label vector whose true label is c; σ is the softmax function, G c () represents a specific domain classifier, G lf () represents the local feature extractor, and Ggf() represents the global feature extractor.

6. The bearing diagnosis method with multi-source sub-domain negative migration inhibition according to any one of claims 1 to 5, characterized in that: The feature extractor includes a global feature extractor and a local feature extractor, the global feature extractor has residual and channel expansion downsampling functions, the local feature extractor includes depthwise separable convolution and convolutional attention, and the convolutional attention includes a channel attention mechanism and a spatial attention mechanism.

7. A bearing diagnosis system with multi-source subdomain negative migration suppression, characterized in that: include: A data acquisition module is used to acquire source domain data and target domain data corresponding to the bearing and perform preprocessing, and acquire the bearing data to be tested for preprocessing; A fault diagnosis model construction module is used to construct a fault diagnosis model including a feature extractor, a domain discriminator, a specific domain classifier, and a local maximum mean difference module, wherein the specific domain classifier includes multiple classifiers; the feature extractor extracts the features of the preprocessed source domain data and the target domain data and inputs them into the domain discriminator, the domain discriminator aligns the global distribution of all domain features through domain adversarial alignment and then inputs them into the specific domain classifier, the specific domain classifier outputs the predicted label of the source domain and the pseudo label of the target domain, the local maximum mean difference module combines the pseudo labels output by all classifiers and the local maximum mean difference to align the subdomains of all domains; the local maximum mean difference module combines the pseudo labels output by all classifiers and the local maximum mean difference to align the subdomains of all domains, specifically: The source domain features obtained after the preprocessed source domain data passes through the specific domain classifier are denoted as Z s , the target domain feature obtained after the preprocessed target domain data passes through the specific domain classifier is Z t ; The true label of the source domain is y s , is the true label of the i-th source domain data, N1 is the number of source domain data; Z t The pseudo label of the target domain obtained after the softmax function operation is is the pseudo label of the jth target domain, and N2 is the number of target domain data; The source domain feature Z s , target domain feature Z t , the true label of each source domain data Pseudo labels for each target domain Input the local maximum mean difference module for subdomain alignment, and calculate the mean square error between the pseudo labels of the target domain output by all classifiers. Close the distribution of pseudo labels and align the subdomain distributions of different source domains; A training module, used for inputting the preprocessed source domain data and target domain data into the fault diagnosis model to train the fault diagnosis model to obtain a trained fault diagnosis model; The fault diagnosis module is used to input the preprocessed bearing data to be tested into the trained fault diagnosis model to obtain the bearing fault diagnosis result.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the bearing diagnosis method for multi-source sub-domain negative migration inhibition as described in any one of claims 1 to 6 is implemented.

9. A bearing diagnostic device with multi-source sub-domain negative migration suppression, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the bearing diagnosis method for inhibiting negative migration of multiple source subdomains as described in any one of claims 1 to 6 is implemented.

Images