A bearing fault diagnosis method based on dual domain adaptive neural network

Through the multi-scale feature extractor and dual tag predictor of dual domain adaptation neural network, combined with the adaptive backpropagation coefficient, the identification problem of bearing fault diagnosis under different working conditions is solved, and the generality and diagnostic accuracy of the model are improved. , achieving more accurate bearing fault diagnosis.

CN118518359BActive Publication Date: 2024-11-19NANJING TECH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411002375.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-19
Estimated Expiration
2044-07-25

AI Technical Summary

Technical Problem

Existing bearing fault diagnosis methods are difficult to identify faults under different working conditions. The traditional transfer learning model is poor in general and has low diagnostic accuracy, especially when the distribution of bearing data sets is large.

Method used

Using a dual domain adaptation neural network method, fault sample features are extracted globally and locally through a multi-scale feature extractor, combined with a dual tag predictor, and reducing edge distribution differences between data sets, and optimizing feature extraction through adaptive backpropagation coefficients The adversarial ability of the device improves the generalization ability and diagnostic accuracy of the model.

Benefits of technology

The fault diagnosis model trained with known data is implemented to apply the bearing fault diagnosis of unknown data, which improves the generality and accuracy of bearing fault diagnosis and outputs more accurate fault diagnosis results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118518359B_ABST
    Figure CN118518359B_ABST
Patent Text Reader

Abstract

The present invention relates to a bearing fault diagnosis method based on a dual domain adaptive neural network, comprising: using a sensor to collect bearing operation data to obtain a source domain data set and a target domain data set; inputting the source domain data set and the target domain data set into a multi-scale feature extractor to obtain mapped source domain samples and target domain samples; based on the source domain samples and the target domain samples, calculating adaptive back propagation coefficients, training a multi-scale feature extractor and a convolutional neural network; based on the trained multi-scale feature extractor, mapping the bearing data to be detected to obtain mapped bearing samples; inputting the mapped bearing samples into a convolutional neural network, and judging the type of bearing fault according to the output label. The present invention solves the problem that the model cannot be trained when the bearing to be detected lacks data, and can output more accurate bearing fault diagnosis results, thereby achieving the purpose of improving the versatility and accuracy of the bearing fault diagnosis model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method based on a dual domain adaptive neural network. Background Art

[0002] With the rapid development of modern industry, the role of rotating machinery in production has become increasingly prominent, and rotary motion has become an indispensable part of modern industrial production. As an important component of rotating machinery, the status and role of rolling bearings have become increasingly prominent. The normality of rolling bearings is directly related to the life and operating status of the machine. When the machine is running at high speed, if the bearing fails, it will cause the machine to run unsteadily or even stop, which will not only lead to a decrease in production efficiency, but also may cause serious safety hazards. Therefore, ensuring the normal operation of rolling bearings is not only a requirement for equipment reliability, but also a responsibility for production safety.

[0003] The successful application of deep learning in machine fault diagnosis is based on the assumption that the training data set and the test data set belong to the same distribution. However, in the actual scenarios of industrial applications, this subordination is difficult to meet: on the one hand, the working conditions of rotating machinery are easy to change; on the other hand, it is difficult to obtain vibration data and labels to train specific models for each working condition. In real production activities, the working conditions of bearings are complex and changeable. The neural network trained based on the data under a single working condition is often difficult to effectively identify faults under other working conditions. However, obtaining a large amount of fault data under all working conditions and building separate deep learning models separately requires a lot of economic and manpower, which is extremely difficult to implement.

[0004] For this type of problem, rolling bearing fault diagnosis based on transfer learning has attracted extensive research in recent years. Transfer learning uses the knowledge of known domains (source domains) to solve problems in related fields (target domains), and has the potential to solve the problems of poor generalization of deep learning models and difficult data collection. However, traditional transfer learning research focuses more on the overall distribution differences between source domain datasets and target domain datasets, while ignoring the distribution differences between samples with the same fault labels. This makes it difficult for traditional models to improve the diagnostic accuracy of bearing faults, although they can reduce the distribution differences between datasets. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a bearing fault diagnosis method based on a dual domain adaptive neural network, which solves the problems of being unable to train a model when the bearing to be detected lacks data, the poor versatility of traditional bearing fault diagnosis methods, and the low diagnostic accuracy of bearing faults. It can output more accurate bearing fault diagnosis results, and can realize the application of the fault diagnosis model trained based on known bearing data to the fault diagnosis of unknown data bearings, thereby achieving the purpose of improving the versatility and accuracy of the bearing fault diagnosis model. This method can extract fault sample features from both global and local aspects through a multi-scale feature extractor, thereby improving the model feature extraction capability and bearing fault diagnosis accuracy; it reduces the edge distribution differences between different bearing data sets through a dual label predictor, thereby improving the generalization ability of the dual domain adaptive neural network; it reasonably controls the adversarial ability of the multi-scale feature extractor through an adaptive back-propagation coefficient, thereby enhancing the training effect of the dual domain adaptive neural network.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a bearing fault diagnosis method based on a dual domain adaptive neural network, comprising the following steps:

[0007] S1. Use sensors to collect bearing operation data and obtain source domain data sets and target domain dataset ;

[0008] S2. Source domain dataset and target domain dataset Input the multi-scale feature extractor respectively to obtain the mapped source domain samples and mapping target domain samples ;

[0009] S3. Mapping source domain samples based on the output of the multi-scale feature extractor and mapping target domain samples , calculate the adaptive back propagation coefficient and ;

[0010] S4. Construct a dual label predictor, including a source domain label predictor module, a domain discriminator module, and a target domain label predictor module, to map the source domain samples and mapping target domain samples Input the three modules respectively, obtain the recognition results and calculate the loss function;

[0011] S5. Based on the calculated adaptive back propagation coefficient and loss function, update all module parameters included in the multi-scale feature extractor and the dual label predictor, and then train the multi-scale feature extractor and the convolutional neural network;

[0012] S6. Based on the trained multi-scale feature extractor, map the bearing operation data sample to be detected to obtain the mapped bearing sample ;

[0013] S7. Mapped bearing samples Input the trained convolutional neural network and determine the type of bearing fault based on the output label.

[0014] Furthermore, in step S1, the bearing operation data is collected by using sensors to obtain the source domain data set and target domain dataset The specific process includes the following steps:

[0015] S11. Install the vibration signal sensor on the bearing and ensure that the sensor is correctly connected and can work normally;

[0016] S12. Collect operating data of different bearings and record them as source domain data set and target domain data set respectively.

[0017] Furthermore, in step S2, the source domain dataset and target domain dataset Input the multi-scale feature extractor respectively to obtain the mapped source domain samples and mapping target domain samples The specific process includes the following steps:

[0018] S21. Establish a multi-scale feature extractor including a self-attention module and a convolution module, and initialize the parameters;

[0019] S22. Source domain dataset and target domain dataset Input the self-attention module and get the output respectively and ; The formula of the self-attention module is as follows:

[0020] ;

[0021] in, express Activation function, represents any sample input into the self-attention module, Representation sample The result output by the self-attention module is: represents the convolution calculation, and Respectively represent the input samples The matrix obtained by convolution calculation;

[0022] S23. Source domain dataset and target domain dataset Input convolution module and get output and ;

[0023] S24. Source domain dataset Output from the self-attention module And the output of the convolution module Merge into , the target domain dataset Output from the self-attention module And the output of the convolution module Merge into ,Will and Integrate them separately to obtain the source domain dataset mapped to the source domain sample output by the multi-feature extractor The target domain samples mapped to the target domain dataset output by the multi-feature extractor ;

[0024] ;

[0025] in, express Activation function, represents the output of the convolution module, represents the output of the self-attention module, represents a sample map output by a multi-scale feature extractor, Represents convolution calculation.

[0026] Further, in step S3, based on the mapped source domain samples output by the multi-scale feature extractor and mapping target domain samples , calculate the adaptive back propagation coefficient and , specifically including: Mapping source domain samples based on the output of the multi-scale feature extractor and mapping target domain samples , calculate the adaptive back propagation coefficient and ;

[0027] ;

[0028] ;

[0029] in, represents the hyperbolic tangent function, represents the mathematical expectation, Indicates taking the absolute value, represents the adaptive back propagation coefficient The upper limit is usually 0.15. represents the adaptive back propagation coefficient The lower limit is generally taken as 0.05.

[0030] Further, in step S5, based on the calculated adaptive back propagation coefficient and loss function, all module parameters included in the multi-scale feature extractor and the dual label predictor are updated, and then the multi-scale feature extractor and the convolutional neural network are trained. The specific process includes the following steps:

[0031] S51, mapping source domain samples and mapping target domain samples Input domain discriminator, get domain prediction result, and calculate loss function ;

[0032] ;

[0033] in, Represents the total number of samples of the input domain discriminator; Represents the domain discriminator for the The predicted domain labels of samples; Indicates The true domain labels of samples;

[0034] S52. Based on loss function Update the domain discriminator parameters based on the loss function Update multi-scale feature extractor parameters;

[0035] S53, mapping source domain samples Input the source domain label predictor, get the label prediction result, and calculate the loss function ;

[0036] ;

[0037] in, is the total number of input source domain samples, represents the prediction result of the source domain label predictor for the source domain sample, Indicates Calculate the cross entropy loss function for each sample;

[0038] S54. Based on loss function Update the multi-scale feature extractor parameters and source domain label predictor parameters;

[0039] S55, mapping the target domain sample Input the source domain label predictor, get the label prediction result, and calculate the loss function ;

[0040] ;

[0041] in, is the total number of input target domain samples, represents the source domain label predictor for the The prediction results of target domain samples are Expressing the Calculate the cross entropy loss function for each sample;

[0042] S56. Based on loss function Update multi-scale feature extractor parameters;

[0043] S57, mapping the target domain sample Input the target domain label predictor, get the label prediction result, and calculate the loss function ;

[0044] ;

[0045] in, is the total number of input target domain samples, represents the target domain label predictor for the The prediction results of target domain samples are Expressing the Calculate the cross entropy loss function for each sample;

[0046] S58. Based on loss function Update the multi-scale feature extractor parameters and target domain label predictor parameters;

[0047] S59, mapping source domain samples Input the target domain label predictor, get the label prediction result, and calculate the loss function ;

[0048] ;

[0049] in, is the total number of input source domain samples, represents the target domain label predictor for the The prediction results of source domain samples are Expressing the Calculate the cross entropy loss function for each sample;

[0050] S510, based on loss function Update multi-scale feature extractor parameters;

[0051] S511, repeating steps S51-S510 to train a dual domain adaptation neural network including a multi-scale feature extractor, a dual label predictor and an adaptive back propagation coefficient;

[0052] S512, inputting the source domain data set into the trained multi-scale feature extractor, and outputting the trained mapped source domain samples;

[0053] S513, inputting the trained mapped source domain samples into the convolutional neural network, calculating the loss and updating the convolutional neural network parameters;

[0054] S514. Repeat steps S512-S513 to train the convolutional neural network.

[0055] Further, in step S6, based on the trained multi-scale feature extractor, the bearing operation data sample to be detected is mapped to obtain the mapped bearing sample The specific process includes the following steps:

[0056] S61. Using the sensor installed on the bearing to be tested, collect the running data sample of the bearing to be tested, recorded as ;

[0057] S62: Input the running data sample of the bearing to be tested into the self-attention module to obtain the output of the self-attention module ;

[0058] ;

[0059] ;

[0060] in, express Activation function, represents the sample input to the self-attention module, Indicates the running data sample of the bearing to be tested Through the output of the self-attention module, represents the convolution calculation, and Respectively represent the running data samples of the bearing to be tested The matrix obtained by convolution calculation;

[0061] S63, the running data sample of the bearing to be tested Input the convolution module and get the output of the convolution module ;

[0062] S64, the output of the self-attention module And the output of the convolution module Merge, and integrate the mapped bearing samples through the trained multi-scale feature extractor ;

[0063] ;

[0064] ;

[0065] in, express Activation function, represents the output of the convolution module, represents the output of the self-attention module, represents the mapped bearing samples obtained by integrating the trained multi-scale feature extractor, Represents convolution calculation.

[0066] Further, in step S7, the mapped bearing samples are Input the trained convolutional neural network and judge the type of bearing fault according to the output label, including: mapping the bearing sample The trained convolutional neural network is input, and the convolutional neural network outputs the probability that the sample belongs to different fault types, and the fault type with the highest probability is taken as the result of rolling bearing fault diagnosis.

[0067] By means of the above technical solution, the present invention provides a bearing fault diagnosis method based on a dual domain adaptive neural network, which has at least the following beneficial effects:

[0068] This method can extract fault sample features from both global and local aspects through a multi-scale feature extractor, thereby improving the accuracy of bearing fault diagnosis; reduce the marginal distribution differences between different bearing data sets through a dual label predictor, thereby improving the generalization ability of the dual domain adaptive neural network; and reasonably control the adversarial ability of the multi-scale feature extractor through an adaptive back-propagation coefficient, thereby enhancing the training effect of the dual domain adaptive neural network. Finally, by designing a dual domain adaptive neural network including a multi-scale feature extractor, a dual label predictor, and an adaptive back-propagation coefficient, the present invention can realize the application of a fault diagnosis model trained based on known bearing data to the fault diagnosis of bearings with unknown data, thereby solving the problem that the model cannot be trained when the bearing to be detected lacks data and the poor versatility of traditional bearing fault diagnosis methods, and can output more accurate bearing fault diagnosis results, thereby achieving the purpose of improving the versatility and accuracy of the bearing fault diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0070] Figure 1 An overall flow chart of a bearing fault diagnosis method based on a dual domain adaptive neural network provided by the present invention;

[0071] Figure 2 An architecture diagram of the multi-scale feature extractor provided by the present invention;

[0072] Figure 3 The architecture diagram of the dual label predictor provided by the present invention;

[0073] Figure 4 It is a schematic diagram of the accuracy of the bearing fault diagnosis method of the present invention;

[0074] Figure 5 A schematic diagram of the loss of the bearing fault diagnosis method of the present invention;

[0075] Figure 6 It is a schematic diagram of the confusion matrix of the bearing fault diagnosis results of the method of the present invention. DETAILED DESCRIPTION

[0076] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods, so that the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0077] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, so the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0078] Please refer to Figure 1-Figure 6 , showing a specific implementation of this embodiment. In this embodiment, the present invention designs a dual domain adaptive neural network including a multi-scale feature extractor, a dual label predictor and an adaptive back propagation coefficient, which can realize the application of the fault diagnosis model trained based on known bearing data to the fault diagnosis of bearings with unknown data, solves the problem that the model cannot be trained when the bearing to be detected lacks data, and can output more accurate bearing fault diagnosis results, thereby achieving the purpose of improving the versatility and accuracy of the bearing fault diagnosis model.

[0079] Please refer to Figure 1This embodiment proposes a bearing fault diagnosis method based on a dual domain adaptive neural network, which includes the following steps:

[0080] S1. Use sensors to collect bearing operation data and obtain source domain data sets and target domain dataset ;

[0081] As a preferred implementation of step S1, in step S1, the bearing operation data is collected by using sensors to obtain the source domain data set. and target domain dataset The specific process includes the following steps:

[0082] S11. Install the vibration signal sensor on the bearing and ensure that the sensor is correctly connected and can work normally;

[0083] S12. Collect operating data of different bearings and record them as source domain data set and target domain data set respectively.

[0084] S2. Source domain dataset and target domain dataset Input the multi-scale feature extractor respectively to obtain the mapped source domain samples and mapping target domain samples ;

[0085] As a preferred implementation of step S2, in step S2, the source domain dataset and target domain dataset Input the multi-scale feature extractor respectively to obtain the mapped source domain samples and mapping target domain samples The specific process includes the following steps:

[0086] S21. Establish a multi-scale feature extractor including a self-attention module and a convolution module, and initialize the parameters;

[0087] S22. Source domain dataset and target domain dataset Input the self-attention module and get the output respectively and ; The formula of the self-attention module is as follows:

[0088] ;

[0089] in, express Activation function, represents any sample input into the self-attention module, Representation sample The result output by the self-attention module is: represents the convolution calculation, and Respectively represent the input samples The matrix obtained by convolution calculation;

[0090] S23. Source domain dataset and target domain dataset Input convolution module and get output and ;

[0091] S24. Source domain dataset Output from the self-attention module And the output of the convolution module Merge into , the target domain dataset Output from the self-attention module And the output of the convolution module Merge into ,Will and Integrate them separately to obtain the source domain dataset mapped to the source domain sample output by the multi-feature extractor The target domain samples mapped to the target domain dataset output by the multi-feature extractor ;

[0092] ;

[0093] in, express Activation function, represents the output of the convolution module, represents the output of the self-attention module, represents a sample map output by a multi-scale feature extractor, Represents convolution calculation.

[0094] In this embodiment, the present invention constructs a multi-scale feature extractor through S2, which can simultaneously extract the global features and local features of the fault signal, enhance the feature extraction capability, and reduce the interference of different working conditions; the multi-scale feature extractor is used to map the bearing data samples, which can extract the common features between different bearing data, thereby solving the problem that the traditional transfer learning method has poor mapping capability and difficulty in extracting common features, such as Figure 2 shown.

[0095] S3. Mapping source domain samples based on the output of the multi-scale feature extractor and mapping target domain samples , calculate the adaptive back propagation coefficient and ;

[0096] As a preferred implementation of step S3, in step S3, based on the mapped source domain samples output by the multi-scale feature extractor and mapping target domain samples , calculate the adaptive back propagation coefficient and , specifically including: Mapping source domain samples based on the output of the multi-scale feature extractor and mapping target domain samples , calculate the adaptive back propagation coefficient and ;

[0097] ;

[0098] ;

[0099] in, represents the hyperbolic tangent function, represents the mathematical expectation, Indicates taking the absolute value, represents the adaptive back propagation coefficient The upper limit is usually 0.15. represents the adaptive back propagation coefficient The lower limit is generally taken as 0.05.

[0100] In this embodiment, by constructing an adaptive back propagation coefficient through S3, the back propagation coefficient can be automatically calculated according to the mapped samples to enhance the model training effect; the loss function is calculated according to the adaptive back propagation coefficient, which can enhance the training effect of the dual domain adaptive neural network, thereby solving the problem that the traditional transfer learning method is difficult to produce confrontation and has poor training effect. Figure 3 shown.

[0101] S4. Construct a dual label predictor, including a source domain label predictor module, a domain discriminator module, and a target domain label predictor module, to map the source domain samples and mapping target domain samples Input the three modules respectively, obtain the recognition results and calculate the loss function;

[0102] S5. Based on the calculated adaptive back propagation coefficient and loss function, update all module parameters included in the multi-scale feature extractor and the dual label predictor, and then train the multi-scale feature extractor and the convolutional neural network;

[0103] As a preferred implementation of step S5, in step S5, based on the calculated adaptive back propagation coefficient and loss function, all module parameters included in the multi-scale feature extractor and the dual label predictor are updated, and then the multi-scale feature extractor and the convolutional neural network are trained. The specific process includes the following steps:

[0104] S51, mapping source domain samples and mapping target domain samples Input domain discriminator, get domain prediction result, and calculate loss function ;

[0105] ;

[0106] in, Represents the total number of samples of the input domain discriminator; Represents the domain discriminator for the The predicted domain labels of samples; Indicates The true domain labels of samples;

[0107] S52. Based on loss function Update the domain discriminator parameters based on the loss function Update multi-scale feature extractor parameters;

[0108] S53, mapping source domain samples Input the source domain label predictor, get the label prediction result, and calculate the loss function ;

[0109] ;

[0110] in, is the total number of input source domain samples, represents the prediction result of the source domain label predictor for the source domain sample, Expressing the Calculate the cross entropy loss function for each sample;

[0111] S54. Based on loss function Update the multi-scale feature extractor parameters and source domain label predictor parameters;

[0112] S55, mapping the target domain sample Input the source domain label predictor, get the label prediction result, and calculate the loss function ;

[0113] ;

[0114] in, is the total number of input target domain samples, represents the source domain label predictor for the The prediction results of target domain samples are Expressing the Calculate the cross entropy loss function for each sample;

[0115] S56. Based on loss function Update multi-scale feature extractor parameters;

[0116] S57, mapping the target domain sample Input the target domain label predictor, get the label prediction result, and calculate the loss function ;

[0117] ;

[0118] in, is the total number of input target domain samples, represents the target domain label predictor for the The prediction results of target domain samples are Expressing the Calculate the cross entropy loss function for each sample;

[0119] S58. Based on loss function Update the multi-scale feature extractor parameters and target domain label predictor parameters;

[0120] S59, mapping source domain samples Input the target domain label predictor, get the label prediction result, and calculate the loss function ;

[0121] ;

[0122] in, is the total number of input source domain samples, represents the target domain label predictor for the The prediction results of source domain samples are Expressing the Calculate the cross entropy loss function for each sample;

[0123] S510, based on loss function Update multi-scale feature extractor parameters;

[0124] S511, repeating steps S51-S510 to train a dual domain adaptation neural network including a multi-scale feature extractor, a dual label predictor and an adaptive back propagation coefficient;

[0125] S512, inputting the source domain data set into the trained multi-scale feature extractor, and outputting the trained mapped source domain samples;

[0126] S513, inputting the trained mapped source domain samples into the convolutional neural network, calculating the loss and updating the convolutional neural network parameters;

[0127] S514. Repeat steps S512-S513 to train the convolutional neural network.

[0128] In this embodiment, the dual label predictor constructed by the present invention in the previous step S4 includes a source domain label predictor, a target domain label predictor and a domain discriminator, which can reduce the marginal distribution differences between different bearing data; updating the multi-scale feature extractor according to the loss function of the dual label predictor can reduce the marginal distribution differences between the same-label data samples, and enhance the feature extraction and mapping capabilities of the multi-scale feature extractor, thereby solving the problem that the traditional transfer learning method cannot reduce the marginal distribution differences between the same-label data samples and the diagnostic accuracy is difficult to improve. Figure 3 shown.

[0129] S6. Based on the trained multi-scale feature extractor, map the bearing operation data sample to be detected to obtain the mapped bearing sample ;

[0130] As a preferred implementation of step S6, in step S6, based on the trained multi-scale feature extractor, the bearing operation data sample to be detected is mapped to obtain the mapped bearing sample The specific process includes the following steps:

[0131] S61. Using the sensor installed on the bearing to be tested, collect the running data sample of the bearing to be tested, recorded as ;

[0132] S62: Input the running data sample of the bearing to be tested into the self-attention module to obtain the output of the self-attention module ;

[0133] ;

[0134] ;

[0135] in, express Activation function, represents the sample input to the self-attention module, Indicates the running data sample of the bearing to be tested Through the output of the self-attention module, represents the convolution calculation, and Respectively represent the running data samples of the bearing to be tested The matrix obtained by convolution calculation;

[0136] S63, the running data sample of the bearing to be tested Input the convolution module and get the output of the convolution module ;

[0137] S64, the output of the self-attention module And the output of the convolution module Merge, and integrate the mapped bearing samples through the trained multi-scale feature extractor ;

[0138] ;

[0139] ;

[0140] in, express Activation function, represents the output of the convolution module, represents the output of the self-attention module, represents the mapped bearing samples obtained by integrating the trained multi-scale feature extractor, Represents convolution calculation.

[0141] S7. Mapped bearing samples Input the trained convolutional neural network and determine the type of bearing fault based on the output label.

[0142] As a preferred implementation of step S7, in step S7, the mapped bearing samples are Input the trained convolutional neural network and judge the type of bearing fault according to the output label, including: mapping the bearing sample The trained convolutional neural network is input, and the convolutional neural network outputs the probability that the sample belongs to different fault types, and the fault type with the highest probability is taken as the result of rolling bearing fault diagnosis.

[0143] More specifically, in order to verify the performance of a bearing fault diagnosis method based on a dual domain adaptive neural network proposed in the present invention, the present method is executed on a public bearing fault data set provided by CWRU and a PT bearing fault data set collected by a laboratory. The CWRU data set includes 40 ".mat" data files, each of which contains a vibration acceleration signal of the bearing collected by an accelerometer placed above the bearing seat at the motor drive end under the condition of a sampling frequency of 12kHz, and all faulty bearings are damaged by single-point electrospark machining. The PT data set contains 4 ".csv" data files, each of which contains a vibration acceleration signal of the bearing collected by an accelerometer placed on the bearing seat at the motor drive end of the PT500mini mechanical fault simulation test bench under the condition of a sampling frequency of 48kHz, and all faulty bearings are damaged by manual cutting. The bearing types collected in the CWRU public bearing fault data set and the PT bearing fault data set are 3 types of faulty bearings and 1 type of normal bearing, and the fault locations of the faulty bearings are the outer ring of the bearing, the rolling element of the bearing, and the inner ring of the bearing. Specifically, Table 1 gives all the bearing fault types and corresponding labels. All data files are sampled, and the signal length of each sampling is 2048. The sampled data sets are divided into training set, validation set and test set. In the experiment, the present invention uses the CWRU public bearing fault data set as the training set and the PT bearing fault data set as the validation set and test set. The training set and the validation set are mapped using a multi-scale feature extractor; the multi-scale feature extractor, the dual label predictor and the convolutional neural network are trained using the mapped training set and the mapped validation set; finally, the trained multi-scale feature extractor is used to map the test set, and the mapped test set is used to verify the fault diagnosis accuracy on the trained convolutional neural network.

[0144] Table 1 Fault types and labels of CWRU dataset and PT dataset

[0145] Damage location normal Inner Circle Rolling element Outer ring Label 0 1 2 3 Training set 48 48 48 48 Validation set 24 24 24 24 Test Set 48 48 48 48

[0146] Figure 2 The following is the architecture diagram of the multi-scale feature extractor. The input data set first passes through the self-attention module and the convolution module to obtain the output of the two modules; the outputs of the two modules are merged into a vector, integrated through the convolution layer, and all elements are converted to the range of 0~1 through the ReLU activation function, and finally the mapping data set output by the multi-scale feature extractor is obtained.

[0147] Figure 3The following is the architecture diagram of the dual label predictor. Based on the mapping dataset samples output by the multi-scale feature extractor, the source domain label predictor, target domain label predictor and domain discriminator in the dual label predictor are input to obtain the recognition results and calculate the loss function respectively; based on the calculated loss function and the adaptive back propagation coefficient, all module parameters contained in the multi-scale feature extractor and the dual label predictor are updated to realize the training of the multi-scale feature extractor and the convolutional neural network.

[0148] As an illustration, Figure 4 and Figure 5 The accuracy and loss changes of the convolutional neural network during the training process are shown. The training of the convolutional neural network uses a mapping data set based on the output of the multi-scale feature extractor that has completed training. The present invention sets the number of training times to 100, the batch size to 64, and the learning rate to an initial 0.001 Adam algorithm. The mapped bearing samples output by the multi-scale feature extractor that has completed training are used as the input of the convolutional neural network. The output of the convolutional neural network is the probability that the input sample belongs to different labels. The bearing fault type corresponding to the label with the highest probability, that is, the bearing fault identification result, can be obtained from Table 1. Figure 4 It can be seen that the training accuracy of the model of the present invention can be higher than 95%; Figure 5 It can be seen that the training loss of the model of the present invention can be lower than 0.1; Figure 6 It can be seen that the model of the present invention has a high diagnostic accuracy rate for all types of faults. Therefore, the bearing fault detection method proposed by the present invention is reliable.

[0149] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0150] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, which can be embodied in any computer-readable medium for use by an instruction execution system, apparatus or device (such as a computer-based system, a system including a processor or other system that can fetch instructions from an instruction execution system, apparatus or device and execute instructions), or used in combination with these instruction execution systems, apparatuses or devices.

[0151] The above implementation methods have been described in detail. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A bearing fault diagnosis method based on dual domain adaptive neural network, characterized in that: The following steps are involved: S1. Use sensors to collect bearing operation data and obtain source domain data sets and target domain dataset ; S2. Source domain dataset and target domain dataset Input the multi-scale feature extractor respectively to obtain the mapped source domain samples and mapping target domain samples ; S3. Mapping source domain samples based on the output of the multi-scale feature extractor and mapping target domain samples , calculate the adaptive back propagation coefficient and ; S4. Construct a dual label predictor, including a source domain label predictor module, a domain discriminator module, and a target domain label predictor module, to map the source domain samples and mapping target domain samples Input the three modules respectively, obtain the recognition results and calculate the loss function; S5. Based on the calculated adaptive back propagation coefficient and loss function, update all module parameters included in the multi-scale feature extractor and the dual label predictor, and then train the multi-scale feature extractor and the convolutional neural network; S6. Based on the trained multi-scale feature extractor, map the bearing operation data sample to be detected to obtain the mapped bearing sample ; S7. Mapped bearing samples Input the trained convolutional neural network and determine the type of bearing fault based on the output label.

2. A bearing fault diagnosis method based on dual domain adaptive neural network according to claim 1, characterized in that: In step S1, the bearing operation data is collected by sensors to obtain the source domain data set and target domain dataset The specific process includes the following steps: S11. Install the vibration signal sensor on the bearing and ensure that the sensor is correctly connected and can work normally; S12. Collect operating data of different bearings and record them as source domain data set and target domain data set respectively.

3. The bearing fault diagnosis method based on dual domain adaptive neural network according to claim 1 is characterized in that: In step S2, the source domain dataset and target domain dataset Input the multi-scale feature extractor respectively to obtain the mapped source domain samples and mapping target domain samples The specific process includes the following steps: S21. Establish a multi-scale feature extractor including a self-attention module and a convolution module, and initialize the parameters; S22. Source domain dataset and target domain dataset Input the self-attention module and get the output respectively and ; The formula of the self-attention module is as follows: ; in, express Activation function, represents any sample input into the self-attention module, Representation sample The result output by the self-attention module is: represents the convolution calculation, and Respectively represent the input samples The matrix obtained by convolution calculation; S23. Source domain dataset and target domain dataset Input convolution module and get output and ; S24. Source domain dataset Output from the self-attention module And the output of the convolution module Merge into , the target domain dataset Output from the self-attention module And the output of the convolution module Merge into ,Will and Integrate them separately to obtain the source domain dataset mapped to the source domain sample output by the multi-feature extractor The target domain samples mapped to the target domain dataset output by the multi-feature extractor ; ; in, express Activation function, represents the output of the convolution module, represents the output of the self-attention module, represents a sample map output by a multi-scale feature extractor, Represents convolution calculation.

4. The bearing fault diagnosis method based on dual domain adaptive neural network according to claim 1 is characterized in that: In step S3, the source domain samples are mapped based on the output of the multi-scale feature extractor and mapping target domain samples , calculate the adaptive back propagation coefficient and , specifically including: Mapping source domain samples based on the output of the multi-scale feature extractor and mapping target domain samples , calculate the adaptive back propagation coefficient and ; ; ; in, represents the hyperbolic tangent function, represents the mathematical expectation, Indicates taking the absolute value, represents the adaptive back propagation coefficient The upper limit of is 0.

15. represents the adaptive back propagation coefficient The lower limit of is 0.

05.

5. The bearing fault diagnosis method based on dual domain adaptive neural network according to claim 1 is characterized in that: In step S5, based on the calculated adaptive back propagation coefficient and loss function, all module parameters included in the multi-scale feature extractor and the dual label predictor are updated, and then the multi-scale feature extractor and the convolutional neural network are trained. The specific process includes the following steps: S51, mapping source domain samples and mapping target domain samples Input domain discriminator, get domain prediction result, and calculate loss function ; ; in, Represents the total number of samples of the input domain discriminator; Represents the domain discriminator for the The predicted domain labels of samples; Indicates The true domain labels of samples; S52. Based on loss function Update the domain discriminator parameters based on the loss function Update multi-scale feature extractor parameters; S53, mapping source domain samples Input the source domain label predictor, get the label prediction result, and calculate the loss function ; ; in, is the total number of input source domain samples, represents the source domain label predictor for the The prediction results of source domain samples are Indicates Calculate the cross entropy loss function for each sample; S54. Based on loss function Update the multi-scale feature extractor parameters and source domain label predictor parameters; S55, mapping the target domain sample Input the source domain label predictor, get the label prediction result, and calculate the loss function ; ; in, is the total number of input target domain samples, represents the source domain label predictor for the The prediction results of target domain samples are Indicates Calculate the cross entropy loss function for each sample; S56. Based on loss function Update multi-scale feature extractor parameters; S57, mapping the target domain sample Input the target domain label predictor, get the label prediction result, and calculate the loss function ; ; in, is the total number of input target domain samples, represents the target domain label predictor for the The prediction results of target domain samples are Indicates Calculate the cross entropy loss function for each sample; S58. Based on loss function Update the multi-scale feature extractor parameters and target domain label predictor parameters; S59, mapping source domain samples Input the target domain label predictor, get the label prediction result, and calculate the loss function ; ; in, is the total number of input source domain samples, represents the target domain label predictor for the The prediction results of source domain samples are Indicates Calculate the cross entropy loss function for each sample; S510, based on loss function Update multi-scale feature extractor parameters; S511, repeating steps S51-S510 to train a dual domain adaptation neural network including a multi-scale feature extractor, a dual label predictor and an adaptive back propagation coefficient; S512, inputting the source domain data set into the trained multi-scale feature extractor, and outputting the trained mapped source domain samples; S513, inputting the trained mapped source domain samples into the convolutional neural network, calculating the loss and updating the convolutional neural network parameters; S514. Repeat steps S512-S513 to train the convolutional neural network.

6. The bearing fault diagnosis method based on dual domain adaptive neural network according to claim 1 is characterized in that: In step S6, based on the trained multi-scale feature extractor, the bearing operation data sample to be detected is mapped to obtain the mapped bearing sample The specific process includes the following steps: S61. Using the sensor installed on the bearing to be tested, collect the running data sample of the bearing to be tested, recorded as ; S62, the running data sample of the bearing to be tested Input the self-attention module and get the output of the self-attention module ; ; ; in, express Activation function, represents the sample input to the self-attention module, Indicates the running data sample of the bearing to be tested Through the output of the self-attention module, represents the convolution calculation, and Respectively represent the running data samples of the bearing to be tested The matrix obtained by convolution calculation; S63, the running data sample of the bearing to be tested Input the convolution module and get the output of the convolution module ; S64, the output of the self-attention module And the output of the convolution module Merge, and integrate the mapped bearing samples through the trained multi-scale feature extractor ; ; ; in, express Activation function, represents the output of the convolution module, represents the output of the self-attention module, represents the mapped bearing samples obtained by integrating the trained multi-scale feature extractor, Represents convolution calculation.

7. The bearing fault diagnosis method based on dual domain adaptive neural network according to claim 1 is characterized in that: In step S7, the mapped bearing samples are Input the trained convolutional neural network and judge the type of bearing fault according to the output label, including: mapping the bearing sample The trained convolutional neural network is input, and the convolutional neural network outputs the probability that the sample belongs to different fault types, and the fault type with the highest probability is taken as the result of rolling bearing fault diagnosis.