A bearing fault diagnosis method based on data selective migration network

Abstract

The application discloses a bearing fault diagnosis method based on a data selective migration network, and comprises the following steps: acquiring 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 data filter to obtain screened source domain samples and target domain samples; inputting the source domain samples and the target domain samples into a multi-scale splicing Unet to obtain mapped source domain samples and mapped target domain samples; training the multi-scale splicing Unet and a dual classifier based on the mapped source domain samples and the mapped target domain samples; mapping bearing operation data samples to be detected based on the trained multi-scale splicing Unet; inputting the mapped bearing samples into the trained target domain classifier to output a bearing fault type judgment result. The application solves the problem that fault features cannot be fully extracted when there is a lack of bearing fault data to be detected, can output more accurate bearing fault diagnosis results, and improves the generality and accuracy of a bearing fault diagnosis model.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a bearing fault diagnosis method based on a data selective migration network. Background Technology

[0002] With the rapid development of modern industry, rotating machinery plays an increasingly prominent role in production, and rotational motion has become an indispensable part of modern industrial production. As a crucial component of rotating machinery, rolling bearings have become increasingly important. The proper functioning of rolling bearings directly affects the lifespan and operating condition of the machine. When a machine is operating at high speed, if a bearing malfunctions, it will lead to unstable operation or even shutdown, which will not only reduce production efficiency but may also 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 and test datasets belong to the same distribution. However, in real-world industrial applications, this relationship is difficult to maintain: on the one hand, neural networks trained on data from a single operating condition often struggle to effectively identify faults under other conditions; on the other hand, acquiring large amounts of fault data from all operating conditions and building separate deep learning models for each requires significant economic and human resources, making implementation extremely challenging.

[0004] For this type of problem, rolling bearing fault diagnosis based on transfer learning has attracted extensive research in recent years. Transfer learning solves problems in related domains (target domains) using known domain knowledge (source domain), and has the potential to address the problems of poor generality and difficulty in data acquisition associated with deep learning models. However, traditional research has not considered that source domain data with huge distribution differences may be useless or have a negative impact on model training, making it difficult for traditional models to effectively improve the accuracy of bearing fault diagnosis. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a bearing fault diagnosis method based on a data selective transfer network (DTL). By using a data filter to rationally select samples with obvious fault features and small distribution differences, the training effect of the DTL is enhanced. Multi-scale concatenation of Unet extracts and integrates multi-scale domain-invariant features of fault samples, improving the model's feature extraction capability and bearing fault diagnosis accuracy. A dual classifier reduces marginal and conditional distribution differences between different bearings, improving the generalization ability of the DTL. This invention aims to solve the problems of insufficient fault feature extraction when bearing fault data is lacking, poor generality of traditional bearing fault diagnosis methods, and low diagnostic accuracy, thereby outputting more accurate bearing fault diagnosis results. It enables the application of a fault diagnosis model trained on known bearing data to the fault diagnosis of bearings with unknown data, thus improving the generality and accuracy of the bearing fault diagnosis model.

[0006] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution:

[0007] A bearing fault diagnosis method based on a data selective migration network specifically includes the following steps:

[0008] S1. Use sensors to collect bearing vibration acceleration signals and obtain source domain datasets. and target domain dataset ;

[0009] S2, Transfer the source domain dataset and target domain dataset Input the data filters separately to select source domain samples from the dataset. and target domain samples ;

[0010] S3, Source Domain Samples and target domain samples Input the multi-scale concatenated Unet separately to obtain the mapped source domain samples. and mapping target domain samples ;

[0011] S4. Construct a dual classifier that includes a source domain classifier and a target domain classifier. Each of the source domain classifier and the target domain classifier contains two fully connected layers, which are respectively input to the mapped source domain samples. and mapping target domain samples Output bearing fault identification results;

[0012] S5. Design a loss function to update the parameter configurations of the multi-scale spliced ​​Unet and the dual classifier until the training of each model is completed;

[0013] S6. Based on the trained multi-scale stitched Unet, map the bearing operation data samples to be detected;

[0014] S7. Map the bearing sample Input the trained target domain classifier and determine the type of bearing fault based on the output label.

[0015] Furthermore, step S1 specifically includes:

[0016] Install the vibration signal sensor onto the bearing, ensuring proper connection and normal operation; collect vibration acceleration signals from the bearing in the laboratory to form a source domain dataset. Collect vibration acceleration signals of the bearing to be tested to form a target domain dataset. .

[0017] Furthermore, step S2 specifically includes:

[0018] S21. Let the i-th sample in the dataset be... It consists of N bearing signals; first calculate the sample Math score The formula is expressed as:

[0019] ;

[0020]

[0021]

[0022] in, Indicates sample The first in One bearing signal, Indicates normalization; Indicates sample The variance of each signal in the data. Indicates sample The mean of each signal;

[0023] S22. Based on the range of values ​​of the signals in the sample, divide the data into 10 equal intervals, and then divide each signal into the corresponding interval based on the magnitude of the bearing vibration acceleration corresponding to each signal.

[0024] S23, Calculate the sample Entropy score The formula is expressed as:

[0025] ;

[0026] ;

[0027] in, For bearing signals discrete probability, express The number of signals contained in the interval. It is a logarithmic function. For the sample Information entropy;

[0028] S24, Calculate the sample Abnormal scores The formula is expressed as:

[0029] ;

[0030] ;

[0031] in, This is the mapping function for the DeepSVDD deep anomaly single-class classification detection model. This indicates that DeepSVDD is applied to the source domain dataset. and target domain dataset The center of the hypersphere in the feature space to be divided; Represents absolute value; Indicates sample The square of the distance to the center of the hypersphere in the characteristic space;

[0032] S25, Record the sample Total score ;

[0033] Then calculate the source domain dataset. and target domain dataset The total score of all samples in the dataset is displayed, and they are sorted from highest to lowest, from the source domain dataset. and target domain dataset The sample with the higher total score is selected as the source domain sample. and target domain samples .

[0034] Furthermore, step S3 specifically includes:

[0035] S31, the multi-scale stitched UNet consists of four max-pooling layers, five gated convolutional modules, and four convolutional modules; regardless of whether it is the source or target domain, the input sample is denoted as... ;

[0036] S32. Four max pooling layers are connected sequentially. Each max pooling layer uses the output of the previous layer as its input, and the input of the first max pooling layer is the input sample. The formula is expressed as:

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] in, The mapping function represents max pooling. , , , These are the pooling features output by the first, second, third, and fourth max-pooling layers, respectively.

[0042] S33. Each gated convolutional module contains three parallel gated convolutional layers with different kernel scales; the first gated convolutional module directly receives the input sample and outputs multi-scale features, expressed by the formula:

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] in, , and These represent the mapping functions corresponding to the three gated convolutional layers, respectively. , and These represent the output features of the three gated convolutional layers in the first gated convolutional module; This represents a mapping function that is superimposed in spatial dimensions; The first gated convolutional module outputs multi-scale features; the second, third, fourth, and fifth gated convolutional modules take the pooling features output from the first to fourth max-pooling layers as input, and after the same processing as the first gated convolutional module, output multi-scale features respectively. , , , ;

[0048] S34. Each convolutional module contains three parallel convolutional layers with different kernel sizes. The first convolutional module takes the multi-scale features output by the fourth and fifth gated convolutional modules as input and outputs integrated features, expressed by the formula:

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] in, Indicates by , The intermediate features obtained through integration; , and These represent the mapping functions corresponding to three convolutional layers with different kernel sizes. , and This represents the output features of three convolutional layers at different scales; The integrated features output by the first convolutional module;

[0055] The second convolutional module takes the integrated features output by the first convolutional module and the multi-scale features output by the third gated convolutional module as input; the third convolutional module takes the integrated features output by the second convolutional module and the multi-scale features output by the second gated convolutional module as input; the fourth convolutional module takes the integrated features output by the third convolutional module and the multi-scale features output by the first gated convolutional module as input; and outputs the integrated features respectively. , , ; and the integrated features output by the fourth convolutional module As the final output of the entire multi-scale spliced ​​Unet ;

[0056] S35. According to steps S31-S34, the source domain samples and target domain samples Input the multi-scale concatenated Unet separately to obtain the mapped source domain samples. and mapping target domain samples .

[0057] Furthermore, step S5 specifically includes:

[0058] S51, Based on mapped source domain samples and mapping target domain samples Calculate the sample difference loss function Update the parameters of the multi-scale stitched Unet, expressed by the formula:

[0059] ;

[0060] ;

[0061] ;

[0062] in, and Represents any two sets of data, This represents the loss function based on the maximum mean difference. This represents the loss function based on Wasserstein distance. Represents data group The total number of samples, Represents data group The total number of samples, Represents data group The One sample, Represents data group The One sample, For kernel functions; respectively Substituting, we get:

[0063] ;

[0064] S52, Map source domain samples Input the source domain classifier and obtain the source domain features output by the first fully connected layer of the source domain classifier. Source domain classification results output by the second fully connected layer ;Map target domain samples Input the target domain classifier and obtain the target domain features output by the first fully connected layer of the target domain classifier. The target domain classification results output by the second fully connected layer ;

[0065] S53. Following the sample difference loss function designed in step S51, based on source domain features... and target domain features Calculate feature loss Based on source domain classification results And target domain classification results Calculate structural loss The formula is expressed as:

[0066] ;

[0067] ;

[0068] Then according to and Update the parameters of the multi-scale stitched Unet, source domain classifier, and target domain classifier;

[0069] S54. Map the target domain samples Input the source domain classifier and obtain the label prediction results of the source domain classifier for the target domain samples. Map source domain samples Input the target domain classifier and obtain the label prediction results of the target domain classifier for the source domain samples. ;

[0070] S55, Based on source domain classification results and source domain real tags Calculate the prediction loss function Update the parameters of the multi-scale stitched UNet and the source domain classifier, as expressed by the formula:

[0071] ;

[0072] ;

[0073] in, This represents the predicted label for any set of samples. This represents the true label corresponding to this group of samples. Indicates the total number of label categories. Indicates label category Recall rate, Define as a tag category The anti-recall rate, Indicates the batch size. Represents the logarithmic function. Indicates the prediction of the first Each sample belongs to the label. The probability, Indicates the first Labels of each sample The numerical value;

[0074] S56. Based on label prediction results and source domain real tags Calculate the loss Update the multi-scale stitched Unet parameters and the target domain classifier parameters; based on the target domain classification results... Label prediction results The label prediction results As pseudo-labels, calculate the loss. Update the parameters of the multi-scale stitched Unet and the target domain classifier;

[0075] S57. Repeat steps S51-S56 until the training of each model is completed.

[0076] Furthermore, step S6 specifically includes:

[0077] S61. Using a sensor installed on the bearing to be tested, collect operating data samples of the bearing to be tested, and record them as follows: ;

[0078] S62. Sample the operating data of the bearing to be tested. Inputting a multi-scale stitched Unet, we obtain the mapped bearing samples. .

[0079] Furthermore, step S7 specifically includes:

[0080] Mapped bearing samples Input the trained target domain classifier, output the probability of the sample belonging to different fault types, and take the fault type with the highest probability as the result of rolling bearing fault diagnosis.

[0081] Based on the above technical solution, the present invention has at least the following beneficial effects:

[0082] This method enhances the training effect of the data selective transfer network by effectively filtering samples with obvious fault features and small distribution differences through a data filter. It improves the model's feature extraction capability and bearing fault diagnosis accuracy by extracting and integrating multi-scale domain-invariant features of fault samples through multi-scale splicing Unet. Furthermore, it improves the generalization ability of the data selective transfer network by reducing marginal and conditional distribution differences among different bearings through a dual classifier. Ultimately, this invention, by designing a data selective transfer network comprising a data filter, multi-scale splicing Unet, and a dual classifier, solves the problems of insufficient fault feature extraction when bearing fault data is lacking, poor generality of traditional bearing fault diagnosis methods, and low diagnostic accuracy. It can output more accurate bearing fault diagnosis results and enables the application of fault diagnosis models trained on known bearing data to the fault diagnosis of bearings with unknown data, thus improving the generality and accuracy of the bearing fault diagnosis model. Attached Figure Description

[0083] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0084] Figure 1 The overall flowchart of a bearing fault diagnosis method based on a data selective migration network provided by the present invention;

[0085] Figure 2 This is an architecture diagram of the multi-scale splicing Unet provided by the present invention;

[0086] Figure 3 This is an architecture diagram of the dual classifier provided by the present invention;

[0087] Figure 4 This diagram illustrates the accuracy of the bearing fault diagnosis method of the present invention.

[0088] Figure 5 This is a schematic diagram illustrating the loss of the bearing fault diagnosis method of the present invention;

[0089] Figure 6 This is a schematic diagram of the confusion matrix of bearing fault diagnosis results obtained by the method of the present invention. Detailed Implementation

[0090] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0091] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this 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 code.

[0092] Please refer to Figures 1-6 This illustration shows a specific implementation of the present invention. In this embodiment, the present invention designs a data selective transfer network that includes a data filter, a multi-scale splicing Unet, and a dual classifier. It solves the problems of insufficient extraction of fault features when there is a lack of bearing fault data to be detected, poor versatility of traditional bearing fault diagnosis methods, and low accuracy of bearing fault diagnosis. It can output more accurate bearing fault diagnosis results and can apply the fault diagnosis model trained based on known bearing data to the fault diagnosis of bearings with unknown data, thereby improving the versatility and accuracy of the bearing fault diagnosis model.

[0093] like Figure 1 As shown, the bearing fault diagnosis method based on data selective migration network proposed in this invention specifically includes the following steps:

[0094] S1. Use sensors to collect bearing vibration acceleration signals and obtain source domain datasets. and target domain dataset ;

[0095] In a preferred embodiment, step S1 specifically comprises:

[0096] Install the vibration signal sensor onto the bearing, ensuring proper connection and normal operation; collect vibration acceleration signals from the bearing in the laboratory to form a source domain dataset. Collect vibration acceleration signals of the bearing to be tested to form a target domain dataset. When a bearing is in operation, the bearing itself will vibrate, and the signal selected in this application is the vibration acceleration value of the bearing at a specific point in time or within a time period.

[0097] S2, Transfer the source domain dataset and target domain dataset Input the data filters separately to select source domain samples from the dataset. and target domain samples ;

[0098] In a preferred embodiment, step S2 specifically includes:

[0099] S21. Let the i-th sample in the dataset be... It consists of N bearing signals; first calculate the sample Math score The formula is expressed as:

[0100] ;

[0101]

[0102]

[0103] in, Indicates sample The first in One bearing signal, Indicates normalization; Indicates sample The variance of each signal in the data. Indicates sample The mean of each signal;

[0104] The larger the mean of the sample signal, the better it reflects the impact of weight changes on the model's diagnostic results; conversely, the closer the mean of the sample signal is to 0, the smaller the impact of weight changes on the diagnostic results; at the same time, the larger the standard deviation of the sample signal, the more obvious the bearing's fault characteristics; conversely, the smaller the standard deviation, the more ambiguous the bearing's fault characteristics.

[0105] This application designs the mathematical score as mean × standard deviation in order to take into account the influence of both the mean and standard deviation when evaluating samples based on the mean and standard deviation. It combines the location and dispersion of the data to reflect the overall trend of the sample data, and tries to select samples with good mean and standard deviation, while avoiding low-quality samples such as those with a mean close to 0 and a very large standard deviation.

[0106] S22. Based on the range of values ​​of the signals in the sample, divide the data into 10 equal intervals, and then divide each signal into the corresponding interval based on the magnitude of the bearing vibration acceleration corresponding to each signal.

[0107] S23, Calculate the sample Entropy score The formula is expressed as:

[0108] ;

[0109]

[0110] ;

[0111] in, For bearing signals discrete probability, express The number of signals contained in the interval. It is a logarithmic function. For the sample Information entropy;

[0112] Information entropy is used to characterize uncertainty; the higher the entropy score, the lower the corresponding information entropy, indicating that the sample signal is more likely to appear in a certain interval, that is, the sample is more certain and the features are more obvious; the lower the entropy score, the higher the corresponding information entropy, indicating that the sample signal tends to be evenly distributed in all intervals, that is, the sample is less certain and the features are relatively vague.

[0113] S24, Calculate the sample Abnormal scores The formula is expressed as:

[0114] ;

[0115] ;

[0116] in, This is the mapping function for the DeepSVDD deep anomaly single-class classification detection model. This indicates that DeepSVDD is applied to the source domain dataset. and target domain dataset The center of the hypersphere in the feature space to be divided; Represents absolute value; Indicates sample The square of the distance to the center of the hypersphere in the characteristic space; The larger the value, the more likely it is to indicate The farther the sample is from the hypercenter of the feature space, the better. The less similar a sample is to the distribution of other samples, the lower its anomaly score is to the distribution of all other samples. In this application, the anomaly score is obtained based on the DeepSVDD anomaly detection model. The lower the anomaly score of a sample, the more similar the sample is to the distribution of all other samples, making training and transfer based on that sample easier and improving the accuracy of fault diagnosis.

[0117] S25, Record the sample Total score In this application, the total score is expressed as a product of three factors to comprehensively evaluate and filter data from three perspectives simultaneously, ultimately selecting samples with excellent scores in all three categories. By using multiplication, extreme cases where some scores are extremely high while others are close to zero can be avoided, thus selecting samples with balanced and excellent scores in all three categories as much as possible.

[0118] Then calculate the source domain dataset. and target domain dataset The total score of all samples in the dataset is displayed, and they are sorted from highest to lowest, from the source domain dataset. and target domain dataset The sample with the higher total score is selected as the source domain sample. and target domain samples .

[0119] S3, Source Domain Samples and target domain samples Input the multi-scale concatenated Unet separately to obtain the mapped source domain samples. and mapping target domain samples ;

[0120] The architecture diagram of the multi-scale spliced ​​Unet constructed in this embodiment is as follows: Figure 2 As shown; however, in a preferred embodiment, step S3 specifically includes:

[0121] S31, the multi-scale stitched UNet consists of four max-pooling layers, five gated convolutional modules, and four convolutional modules; regardless of whether it is the source or target domain, the input sample is denoted as... ;

[0122] S32. Four max pooling layers are connected sequentially. Each max pooling layer uses the output of the previous layer as its input, and the input of the first max pooling layer is the input sample. The formula is expressed as:

[0123] ;

[0124] ;

[0125] ;

[0126] ;

[0127] in, The mapping function represents max pooling. , , , These are the pooling features output by the first, second, third, and fourth max-pooling layers, respectively.

[0128] S33. Each gated convolutional module contains three parallel gated convolutional layers with different kernel scales; the first gated convolutional module directly receives the input sample and outputs multi-scale features, expressed by the formula:

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] in, , and These represent the mapping functions corresponding to the three gated convolutional layers, respectively. , and These represent the output features of the three gated convolutional layers in the first gated convolutional module; This represents a mapping function that is superimposed in spatial dimensions; The first gated convolutional module outputs multi-scale features; the second, third, fourth, and fifth gated convolutional modules take the pooling features output from the first to fourth max-pooling layers as input, and after the same processing as the first gated convolutional module, output multi-scale features respectively. , , , ;

[0134] It should be noted that in this embodiment, the kernel sizes of the three convolutional layers in the same gated convolutional module are different, while the kernel sizes of different gated convolutional modules are the same. For example, assuming that the kernel sizes of the three convolutional layers in the same gated convolutional module are 1×3, 1×5, and 1×7, then the kernel sizes contained in different gated convolutional modules are also 1×3, 1×5, and 1×7, respectively.

[0135] The calculation process of the multi-scale features output by each gated convolutional module refers to the first gated convolutional module, as shown below:

[0136] The multi-scale features output by the second gated convolutional module :

[0137] ;

[0138] ;

[0139] ;

[0140] ;

[0141] The multi-scale features output by the third gated convolutional module :

[0142] ;

[0143] ;

[0144] ;

[0145] ;

[0146] The multi-scale features output by the fourth gated convolutional module :

[0147] ;

[0148] ;

[0149] ;

[0150] ;

[0151] The multi-scale features output by the fifth gated convolutional module :

[0152] ;

[0153] ;

[0154] ;

[0155] ;

[0156] S34. Each convolutional module contains three parallel convolutional layers with different kernel sizes. The first convolutional module takes the multi-scale features output by the fourth and fifth gated convolutional modules as input and outputs integrated features, expressed by the formula:

[0157] ;

[0158] ;

[0159] ;

[0160] ;

[0161] ;

[0162] in, Indicates by , The intermediate features obtained through integration; , and These represent the mapping functions corresponding to three convolutional layers with different kernel sizes. , and This represents the output features of three convolutional layers at different scales; The integrated features output by the first convolutional module;

[0163] The second convolutional module takes the integrated features output by the first convolutional module and the multi-scale features output by the third gated convolutional module as input; the third convolutional module takes the integrated features output by the second convolutional module and the multi-scale features output by the second gated convolutional module as input; the fourth convolutional module takes the integrated features output by the third convolutional module and the multi-scale features output by the first gated convolutional module as input; and outputs the integrated features respectively. , , ; and the integrated features output by the fourth convolutional module As the final output of the entire multi-scale spliced ​​Unet ;

[0164] Similarly, the kernel scales of different convolutional modules are the same; the calculation process of the integrated features output by each convolutional module is also as follows:

[0165] The integrated features output by the second convolutional module :

[0166] ;

[0167] ;

[0168] ;

[0169] ;

[0170] ;

[0171] The integrated features output by the third convolutional module :

[0172] ;

[0173] ;

[0174] ;

[0175] ;

[0176] ;

[0177] The integrated features output by the fourth convolutional module :

[0178] ;

[0179] ;

[0180] ;

[0181] ;

[0182] ;

[0183] S35. According to steps S31-S34, the source domain samples and target domain samples Input the multi-scale concatenated Unet separately to obtain the mapped source domain samples. and mapping target domain samples .

[0184] Compared to traditional models, the multi-scale stitched Unet designed in this application incorporates a gated convolutional module with a gating mechanism that dynamically adjusts the activation level of the convolutional kernel based on input features, selectively focusing on features in different contexts. The convolutional module uses parallel convolutional layers with different kernel sizes to adaptively extract and integrate implicit features of different scales from fault signals. This multi-scale stitched Unet enables the extraction and integration of multi-scale domain-invariant features from fault samples, thereby improving the model's feature extraction capability and bearing fault diagnosis accuracy.

[0185] S4. Construct a dual classifier that includes a source domain classifier and a target domain classifier. Each of the source domain classifier and the target domain classifier contains two fully connected layers, which are respectively input to the mapped source domain samples. and mapping target domain samples The output shows the bearing fault identification results; the architecture diagram of the dual classifier in this embodiment is as follows. Figure 3 As shown;

[0186] S5. Design a loss function to update the parameter configurations of the multi-scale spliced ​​Unet and the dual classifier until the training of each model is completed;

[0187] In a preferred embodiment, step S5 specifically includes:

[0188] S51, Based on mapped source domain samples and mapping target domain samples Calculate the sample difference loss function Update the parameters of the multi-scale stitched Unet, expressed by the formula:

[0189] ;

[0190] ;

[0191] ;

[0192] in, and Represents any two sets of data, This represents the loss function based on the maximum mean difference. This represents the loss function based on Wasserstein distance. Represents data group The total number of samples, Represents data group The total number of samples, Represents data group The One sample, Represents data group The One sample, For kernel functions; respectively Substituting, we get:

[0193] ;

[0194] S52, Map source domain samples Input the source domain classifier and obtain the source domain features output by the first fully connected layer of the source domain classifier. Source domain classification results output by the second fully connected layer ;Map target domain samples Input the target domain classifier and obtain the target domain features output by the first fully connected layer of the target domain classifier. The target domain classification results output by the second fully connected layer ;

[0195] S53. Following the sample difference loss function designed in step S51, based on source domain features... and target domain features Calculate feature loss Based on source domain classification results And target domain classification results Calculate structural loss The formula is expressed as:

[0196] ;

[0197] ;

[0198] Then according to and Update the parameters of the multi-scale stitched Unet, source domain classifier, and target domain classifier;

[0199] S54. Map the target domain samples Input the source domain classifier and obtain the label prediction results of the source domain classifier for the target domain samples. Map source domain samples Input the target domain classifier and obtain the label prediction results of the target domain classifier for the source domain samples. ;

[0200] S55, Based on source domain classification results and source domain real tags Calculate the prediction loss function Update the parameters of the multi-scale stitched UNet and the source domain classifier, as expressed by the formula:

[0201] ;

[0202]

[0203] in, This represents the predicted label for any set of samples. This represents the true label corresponding to this group of samples. Indicates the total number of label categories. Indicates label category Recall rate, Define as a tag category The anti-recall rate, Indicates the batch size. Represents the logarithmic function. Indicates the prediction of the first Each sample belongs to the label. The probability, Indicates the first Labels of each sample The numerical value;

[0204] S56. Based on label prediction results and source domain real tags Calculate the loss Update the multi-scale stitched Unet parameters and the target domain classifier parameters; based on the target domain classification results... Label prediction results The label prediction results As pseudo-labels, calculate the loss. Update the parameters of the multi-scale stitched Unet and the target domain classifier;

[0205] S57. Repeat steps S51-S56 until the training of each model is completed.

[0206] It should be noted here that the loss function designed in this application... and The parameters of the multi-scale stitched Unet will be updated. Based on this, It will update the parameters of both the source domain classifier and the target domain classifier. It selectively updates the parameters of the source domain classifier and the target domain classifier based on different input data. For example... Figure 3 As shown, the sources of the red and yellow dashed arrows represent different Ls and Lc, and the direction of the arrows indicates which part of the parameters the loss function will update.

[0207] The two are calculated and used in completely different ways. First, The calculation is based on and The sum of squares; The calculation is based on the recall rate of a batch of samples multiplied by the cross-entropy. Secondly, This represents the difference in distribution between any two samples, used to express the degree of similarity between two batches of samples. The larger the value, the smaller the similarity between samples; This represents the difference between the predicted label and the true label, indicating the accuracy of the model in identifying the fault type of a batch of samples. The larger the value, the lower the classification accuracy of the model.

[0208] S6. Based on the trained multi-scale stitched Unet, map the bearing operation data samples to be detected;

[0209] In a preferred embodiment, step S6 specifically includes:

[0210] S61. Using a sensor installed on the bearing to be tested, collect operating data samples of the bearing to be tested, and record them as follows: ;

[0211] S62. Sample the operating data of the bearing to be tested. Inputting a multi-scale stitched Unet, we obtain the mapped bearing samples. .

[0212] S7. Map the bearing sample Input the trained target domain classifier and determine the type of bearing fault based on the output label;

[0213] In a preferred embodiment, step S7 specifically includes:

[0214] Mapped bearing samples Input the trained target domain classifier, output the probability of the sample belonging to different fault types, and take the fault type with the highest probability as the result of rolling bearing fault diagnosis.

[0215] Thus, the method proposed in this invention comprises a data selective transfer network including a data filter, a multi-scale splicing Unet, and a dual classifier. To verify the performance of the proposed method, it was executed on the publicly available bearing fault dataset provided by CWRU and the PT bearing fault dataset collected in the laboratory. The CWRU dataset contains bearing vibration acceleration signals acquired by placing an accelerometer above the bearing housing at the motor drive end under a sampling frequency of 12kHz. All faulty bearings in the CWRU dataset exhibited single-point damage from electrical discharge machining. The PT dataset contains bearing vibration acceleration signals acquired by placing an accelerometer on the bearing housing at the motor drive end of a PT500mini mechanical fault simulation test bench under a sampling frequency of 48kHz. All faulty bearings in the PT dataset exhibited damage from manual cutting. The bearing types collected in the CWRU and PT bearing fault datasets include three types of faulty bearings and one type of normal bearing. The fault locations of the faulty bearings are the outer ring, rolling elements, and inner ring, with fault diameters of 0.18, 0.36, and 0.54 mm, respectively. Specifically, Table 1 below lists all bearing failure types and their corresponding labels:

[0216] All data files are sampled, with each sample having a signal length of 1024.

[0217] Table 1. Fault types and labels in the CWRU and PT datasets

[0218]

[0219] In the experiment, this invention uses the CWRU public bearing fault dataset as the training set and the PT bearing fault dataset as the validation and test sets. The multi-scale stitched Unet is used to map the training and validation sets; the multi-scale stitched Unet, source domain classifier, and target domain classifier are trained using the mapped training and validation sets; finally, the trained multi-scale stitched Unet is used to map the test set, and the fault diagnosis accuracy is verified on the trained target domain classifier using the mapped test set.

[0220] As an explanation, Figure 4 and Figure 5 This paper demonstrates the accuracy and loss changes of the data selective transfer network during training. The training iterations were set to 100, the batch size to 64, and the learning rate to an initial Adam algorithm of 0.001. The mapped bearing samples output from the multi-scale stitched Unet after training were used as input to the target domain classifier. The output of the target domain classifier is the probability that the input sample belongs to different labels. Table 1 shows the bearing fault type corresponding to the label with the highest probability, i.e., the bearing fault identification result. Figure 4It can be seen that the training accuracy of the model of this invention can be higher than 94%; Figure 5 It can be seen that the training loss of the model in this invention can be lower than 0.2; from Figure 6 As can be seen, the model of this invention has a high diagnostic accuracy for all types of faults. Therefore, the bearing fault method proposed in this invention is reliable.

[0221] In summary, the method proposed in this invention solves the problems of insufficient extraction of fault features when there is a lack of bearing fault data, poor versatility of traditional bearing fault diagnosis methods, and low accuracy of bearing fault diagnosis. It can output more accurate bearing fault diagnosis results and can apply the fault diagnosis model trained based on known bearing data to the fault diagnosis of bearings with unknown data, thereby improving the versatility and accuracy of the bearing fault diagnosis model.

[0222] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0223] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0224] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A bearing fault diagnosis method based on a data selective migration network, characterized in that, Specifically, the following steps are included: S1. Use sensors to collect bearing vibration acceleration signals and obtain source domain datasets. and target domain dataset ; S2, Transfer the source domain dataset and target domain dataset Input the data filters separately to select source domain samples from the dataset. and target domain samples ; Step S2 specifically includes: S21. Let the i-th sample in the dataset be... It consists of N bearing signals; first calculate the sample Math score The formula is expressed as: ； ； ； in, Indicates sample The first in One bearing signal, Indicates normalization; Indicates sample The variance of each signal in the data. Indicates sample The mean of each signal; S22. Based on the range of values ​​of the signals in the sample, divide the data into 10 equal intervals, and then divide each signal into the corresponding interval based on the magnitude of the bearing vibration acceleration corresponding to each signal. S23, Calculate the sample Entropy score The formula is expressed as: ； ； ； in, For bearing signals discrete probability, express The number of signals contained in the interval. It is a logarithmic function. For the sample Information entropy; S24, Calculate the sample Abnormal scores The formula is expressed as: ； ； in, This is the mapping function for the DeepSVDD deep anomaly single-class classification detection model. This indicates that DeepSVDD is applied to the source domain dataset. and target domain dataset The center of the hypersphere in the feature space to be divided; Represents absolute value; Indicates sample The square of the distance to the center of the hypersphere in the characteristic space; S25, Record the sample Total score ; Then calculate the source domain dataset. and target domain dataset The total score of all samples in the dataset is sorted from highest to lowest, and the data is from the source domain dataset. and target domain dataset The sample with the higher total score is selected as the source domain sample. and target domain samples ; S3, Source Domain Samples and target domain samples Input the multi-scale concatenated Unet separately to obtain the mapped source domain samples. and mapping target domain samples ; S4. Construct a dual classifier that includes a source domain classifier and a target domain classifier. Each of the source domain classifier and the target domain classifier contains two fully connected layers, which are respectively input to the mapped source domain samples. and mapping target domain samples Output bearing fault identification results; S5. Design a loss function to update the parameter configurations of the multi-scale spliced ​​Unet and the dual classifier until the training of each model is completed; S6. Based on the trained multi-scale stitched Unet, map the bearing operation data samples to be detected; S7. Map the bearing sample Input the trained target domain classifier and determine the type of bearing fault based on the output label.

2. The bearing fault diagnosis method based on a data selective migration network according to claim 1, characterized in that, Step S1 is as follows: Install the vibration signal sensor onto the bearing, ensuring the sensor is correctly connected and functioning properly; collect vibration acceleration signals from the bearing in the laboratory to form a source domain dataset. Collect vibration acceleration signals of the bearing to be tested to form a target domain dataset. .

3. The bearing fault diagnosis method based on a data selective migration network according to claim 1, characterized in that, Step S3 specifically includes: S31, the multi-scale stitched UNet consists of four max-pooling layers, five gated convolutional modules, and four convolutional modules; regardless of whether it is the source or target domain, the input sample is denoted as... ; S32. Four max pooling layers are connected sequentially. Each max pooling layer uses the output of the previous layer as its input, and the input of the first max pooling layer is the input sample. The formula is expressed as: ； ； ； ； in, The mapping function represents max pooling. , , , These are the pooling features output by the first, second, third, and fourth max-pooling layers, respectively. S33. Each gated convolutional module contains three parallel gated convolutional layers with different kernel scales; the first gated convolutional module directly receives the input sample and outputs multi-scale features, expressed by the formula: ； ； ； ； in, , and These represent the mapping functions corresponding to the three gated convolutional layers, respectively. , and These represent the output features of the three gated convolutional layers in the first gated convolutional module; This represents a mapping function that is superimposed in spatial dimensions; The first gated convolutional module outputs multi-scale features; the second, third, fourth, and fifth gated convolutional modules take the pooling features output from the first to fourth max-pooling layers as input, and after the same processing as the first gated convolutional module, output multi-scale features respectively. , , , ; S34. Each convolutional module contains three parallel convolutional layers with different kernel sizes. The first convolutional module takes the multi-scale features output by the fourth and fifth gated convolutional modules as input and outputs integrated features, expressed by the formula: ； ； ； ； ； in, Indicates by , Intermediate features obtained through integration; , and These represent the mapping functions corresponding to three convolutional layers with different kernel sizes. , and This represents the output features of three convolutional layers at different scales; The integrated features output by the first convolutional module; The second convolutional module takes the integrated features output by the first convolutional module and the multi-scale features output by the third gated convolutional module as input; the third convolutional module takes the integrated features output by the second convolutional module and the multi-scale features output by the second gated convolutional module as input; the fourth convolutional module takes the integrated features output by the third convolutional module and the multi-scale features output by the first gated convolutional module as input; and outputs the integrated features respectively. , , ; and the integrated features output by the fourth convolutional module As the final output of the entire multi-scale spliced ​​Unet ; S35. According to steps S31-S34, the source domain samples and target domain samples Input the multi-scale concatenated Unet separately to obtain the mapped source domain samples. and mapping target domain samples .

4. The bearing fault diagnosis method based on a data selective migration network according to claim 1, characterized in that, Step S5 specifically includes: S51, Based on mapped source domain samples and mapping target domain samples Calculate the sample difference loss function Update the parameters of the multi-scale stitched Unet, expressed by the formula: ； ； ； in, and Represents any two sets of data, This represents the loss function based on the maximum mean difference. This represents the loss function based on Wasserstein distance. Represents data group The total number of samples, Represents data group The total number of samples, Represents data group The One sample, Represents data group The One sample, For kernel functions; respectively Substituting, we get: ； S52, Map source domain samples Input the source domain classifier and obtain the source domain features output by the first fully connected layer of the source domain classifier. Source domain classification results output by the second fully connected layer Map the target domain samples Input the target domain classifier and obtain the target domain features output by the first fully connected layer of the target domain classifier. The target domain classification results output by the second fully connected layer ; S53. Following the sample difference loss function designed in step S51, based on source domain features... and target domain features Calculate feature loss Based on source domain classification results And target domain classification results Calculate structural loss The formula is expressed as: ； ； Then according to and Update the parameters of the multi-scale stitched Unet, source domain classifier, and target domain classifier; S54. Map the target domain samples Input the source domain classifier and obtain the label prediction results of the source domain classifier for the target domain samples. Map source domain samples Input the target domain classifier and obtain the label prediction results of the target domain classifier for the source domain samples. ; S55, Based on source domain classification results and source domain real tags Calculate the prediction loss function Update the parameters of the multi-scale stitched UNet and the source domain classifier, as expressed by the formula: ； ； in, This represents the predicted label for any set of samples. This represents the true label corresponding to this group of samples. Indicates the total number of label categories. Indicates label category Recall rate, Define as a tag category The anti-recall rate, Indicates the batch size. Represents the logarithmic function. Indicates the prediction of the first Each sample belongs to the label. The probability, Indicates the first Labels of each sample The numerical value; S56. Based on label prediction results and source domain real tags Calculate the loss Update the multi-scale stitched Unet parameters and the target domain classifier parameters; based on the target domain classification results... Label prediction results The label prediction results As pseudo-labels, calculate the loss. Update the parameters of the multi-scale stitched Unet and the target domain classifier; S57. Repeat steps S51-S56 until the training of each model is completed.

5. The bearing fault diagnosis method based on a data selective migration network according to claim 1, characterized in that, Step S6 specifically includes: S61. Using a sensor installed on the bearing to be tested, collect operating data samples of the bearing to be tested, and record them as follows: ; S62. Sample the operating data of the bearing to be tested. Inputting a multi-scale stitched Unet, we obtain the mapped bearing samples. .

6. The bearing fault diagnosis method based on a data selective migration network according to claim 1, characterized in that, Step S7 is as follows: Mapped bearing samples Input the trained target domain classifier, output the probability of the sample belonging to different fault types, and take the fault type with the highest probability as the result of rolling bearing fault diagnosis.

Images