Rolling bearing fault diagnosis method based on joint weighted multi-scale migration algorithm

Abstract

A rolling bearing fault diagnosis method based on a joint weighted multi-scale migration algorithm is used for enhancing confusion and alignment of cross-domain features through a distribution loss measurement method based on JSD, so that the influence caused by inconsistent distribution of two-domain data fault features is reduced; then, aiming at the negative migration problem possibly occurring in the feature extraction process, designing an improved pseudo-label target domain classification loss mechanism, effectively adjusting loss weights of different target domain samples through adaptive weight distribution, inhibiting a negative migration effect, and promoting a feature extractor to learn more robust and cross-domain consistent features; and finally, through a JWMS-NET combined weighted multi-scale migration neural network, feature representation of a target domain can be optimized, a decision boundary is flexibly controlled, and the diagnosis precision is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearing fault diagnosis, and in particular to a rolling bearing fault diagnosis method based on a joint weighted multi-scale migration algorithm. Background Art

[0002] With the great progress of information technology, the data-driven neural network fault diagnosis method has developed very rapidly, and many excellent fault diagnosis models have emerged. These fault diagnosis technologies use a large amount of labeled fault data for network training to enable them to learn fault features, so as to achieve good fault classification. The whole process does not require manual operation and calculation, greatly improving the fault diagnosis effect. However, this type of method usually relies on a large-scale labeled data set for training and assumes that the training data in the source domain and the target domain have the same distribution. Due to these data conditions, the development of such fault diagnosis methods is restricted. Therefore, how to design a highly adaptable fault diagnosis method in the case of lack of labeled data and data distribution differences has become an important research direction. Facing the above problems, the fault diagnosis industry has conducted a large number of experimental studies and compared with other models, and the fault diagnosis model based on domain adaptation and transfer learning has been gradually developed and achieved good results in practical engineering applications.

[0003] Although cross-condition transfer learning has made some progress in the fault diagnosis of rotating machinery in recent years, existing research still faces multiple challenges, especially in the handling of the feature distribution differences and negative transfer phenomena between the two domains. The problems mainly include: (1) The distribution alignment problem between the source domain and the target domain: The feature distribution differences between the two domains are not effectively aligned. Simple alignment techniques can only handle simple categories, but for difficult categories, the transfer effect is poor. (2) The negative transfer problem: In the scenario where there is a lack of labeled data in the target domain, the feature extractor is too dependent on the source domain data, resulting in overfitting or offset problems, thus presenting the negative transfer phenomenon. Summary of the Invention

[0004] The purpose of the present invention is to provide a rolling bearing fault diagnosis method based on a joint weighted multi-scale migration algorithm, which reduces the influence caused by the inconsistent fault feature distribution of the data in the two domains and improves the diagnosis accuracy.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: A rolling bearing fault diagnosis method based on a joint weighted multi-scale migration algorithm, comprising the following steps: Step S1, obtaining vibration signals of a rolling bearing under multiple different rotational speed conditions through a public data set, and segmenting the vibration signals through a data overlapping segmentation method to obtain a source domain sample data set and a target domain sample data set; Step S2: input the source domain sample dataset and the target domain sample dataset into the multi-scale dynamic convolutional residual network (MDCRN) to obtain the source domain fault feature latent vector and the target domain fault feature latent vector; Step S3: Input the source domain fault feature latent vector and the target domain fault feature latent vector into the domain adaptation module, and convert the extracted features of the source domain and target domain into probability distribution, and then use the JSD loss function to calculate the source domain and target domain. JSD Distribution distance value L jsd ( P , Q ); The JSD loss function is: (1); In formula (1), P and Q are the feature distributions of the source domain respectively X s and the feature distribution of the target domain X t , M = P + Q / 2 , for Q Relative to M The KL divergence of Q arrive M differences; Step S4: Input the source domain fault feature potential vector into the classifier. The classifier marks part of the sample data as the true label and marks the rest of the sample data as the predicted label through the pseudo label method. Then, the source domain cross entropy loss function is used to calculate the error value between the true label and the predicted label. L ( X s , Y s ), as the source domain cross entropy loss L x ; The source domain cross entropy loss function is: (2); In formula (2), E represents the mathematical expectation, x j s represents the true label, y j s represents the predicted label, f cross （f） represents the cross entropy loss function; Step S5: Input the potential vector of the target domain fault features into the classifier. The classifier marks some sample data as true labels and marks the remaining sample data as predicted labels using the pseudo-label method. Then, calculate the error value between the true labels and the predicted labels using the target domain cross-entropy loss function L ( X t , Y t ), which is used as the target domain cross-entropy loss value L y ; The target domain cross-entropy loss function is: (3); In formula (3), E represents the mathematical expectation, x j t represents the true label, y j t represents the predicted label, f cross （f） represents the cross-entropy loss function; Step S6: Calculate the total loss value of the sample data set using the total loss function L all ; The total loss function is: (4); Then, use the backpropagation algorithm to update the parameters of the feature extractor G f and the classifier G s . After each update, recalculate the total loss value L all , until the total loss value L all is 0 and stop the update; The parameter update method of the feature extractor G f is: (5); The parameter update method of the classifier G s is: (6); In formulas (5) and (6), h represents the learning rate; Step S7: Through the updated feature extractor G f and the classifierG s , and a domain adaptation module is cooperated to construct a JWMS-NET joint weighted multi-scale transfer network. The target domain fault feature latent vector is input into the JWMS-NET neural network, and the discriminated bearing state category of the target domain is output. The discriminated bearing state category of the target domain is compared and calculated with the sample data label of the target domain to obtain the diagnostic recognition accuracy of the rolling bearing fault.

[0006] Preferably, in step S1, the vibration signals of the rolling bearing at different rotational speeds are obtained through the WT-planetary-gearbox-dataset public dataset. The rolling bearing includes four states and operates under four working conditions. The four states include healthy rolling gear, gear damage, tooth root fracture, and missing teeth. The four working conditions include 20Hz, 30Hz, 40Hz, and 50Hz.

[0007] Preferably, in step S1, the vibration signals of the rolling bearing at different rotational speeds are obtained through the bearing dataset of Huazhong University of Science and Technology. The rolling bearing includes nine states and operates under four working conditions. Four selected states include normal, severe inner ring fault, severe rolling element fault, and severe compound fault. The four working conditions include 65Hz, 70Hz, 75Hz, and 80Hz.

[0008] Preferably, in step S1, the vibration signals of the rolling bearing at different rotational speeds are obtained through the bearing fault dataset of Southeast University. The rolling bearing includes five states and operates under two working conditions. Four selected states include roller fault, inner ring fault, compound fault, and normal operation. The two working conditions include 20Hz and 30Hz.

[0009] Preferably, the multi-scale dynamic convolutional residual network MDCRN in step S2 adopts a multi-scale convolutional neural network MSCNN architecture including three parallel branches. Each branch has a dynamic convolutional residual module DC-ResBlock. The dynamic convolutional residual module DC-ResBlock includes the main path of the residual block and the branch path of the residual connection. The main path and the branch path respectively use dynamic convolution for feature extraction, and then fuse dynamic convolution and residual connection.

[0010] Preferably, at each update in step S6, first update the target domain cross-entropy loss value L y through the weight coefficient η, and then calculate the total loss value L' y using the updated L all . L' y The calculation formula of L' y= η· L y (7); (8); In Equation (8), w i = 0.001, w f = 0.1, e represents the current training epoch, e t represents the total number of training epochs.

[0011] According to the above technical solution, the beneficial effects of the present invention are as follows: The present invention uses a distribution loss metric method based on JSD to enhance the confusion and alignment of cross-domain features, thereby reducing the impact caused by the inconsistent distribution of fault features in the two domains of data; then, aiming at the possible negative transfer problem in the feature extraction process, an improved pseudo-label target domain classification loss mechanism is designed, and through adaptive weight allocation, the loss weights of different target domain samples are effectively adjusted, the negative transfer effect is suppressed, and the feature extractor is prompted to learn more robust and cross-domain consistent features; finally, through the JWMS-NET joint weighted multi-scale transfer neural network, the feature representation of the target domain can be optimized, and the decision boundary can be flexibly controlled, thereby further improving the diagnostic accuracy. Description of the Drawings

[0012] Figure 1 is the JWMS-Net structure diagram of the present invention; Figure 2 is the MDCRN feature extractor structure diagram of the present invention; Figure 3 is the dynamic convolutional residual module structure diagram of the present invention; Figure 4 is the flow chart of the present invention. Detailed Embodiments

[0013] Referring to the drawings, the detailed embodiments are as follows: A rolling bearing fault diagnosis method based on a joint weighted multi-scale transfer algorithm, establishing a rotating machinery cross-condition fault diagnosis model based on the joint weighted multi-scale transfer algorithm, the model combines a multi-scale dynamic convolutional residual network MDCRN as a feature extractor, a pseudo-label target domain classification loss, and a JSD distribution loss, specifically including the following steps: Step S1, obtaining vibration signals of a rolling bearing under multiple different rotational speed conditions through a public dataset, and segmenting the vibration signals by a data overlapping segmentation method to obtain a source domain sample dataset and a target domain sample dataset.

[0014] In this embodiment, vibration signals of rolling bearings at different rotational speeds are obtained from the WT-planetary-gearbox-dataset, the bearing dataset of Huazhong University of Science and Technology (HUST), and the bearing fault dataset of Southeast University respectively. Data samples of different datasets are preprocessed, and each sample contains 3072 data points. The collected time-domain data is converted into frequency-domain data through FFT, and the frequency-domain data is used as the input data of the fault diagnosis model. Datasets Ⅰ, Ⅱ, and Ⅲ are constructed respectively, and three cross-condition transfer learning tasks are constructed using the three bearing datasets, as shown in Table 1.

[0015] Table 1 Details of Datasets Ⅰ, Ⅱ, and Ⅲ Specifically, vibration signals of rolling bearings at different rotational speeds are obtained from the publicly available WT-planetary-gearbox-dataset. The rolling bearing has four states and operates under four working conditions. The four states include healthy rolling gear, gear damage, tooth root fracture, and missing teeth. The four working conditions include 20 Hz, 30 Hz, 40 Hz, and 50 Hz. Vibration signals of rolling bearings at different rotational speeds are obtained from the bearing dataset of Huazhong University of Science and Technology. The rolling bearing has nine states and operates under four working conditions. Four selected states include normal, severe inner ring fault, severe rolling element fault, and severe compound fault. The four working conditions include 65 Hz, 70 Hz, 75 Hz, and 80 Hz. Vibration signals of rolling bearings at different rotational speeds are obtained from the bearing fault dataset of Southeast University. The rolling bearing has five states and operates under two working conditions. Four selected states include roller fault, inner ring fault, compound fault, and normal operation. The two working conditions include 20 Hz and 30 Hz.

[0016] Step S2: Input the source domain sample dataset and the target domain sample dataset into the multi-scale dynamic convolutional residual network MDCRN respectively to obtain the source domain fault feature latent vector and the target domain fault feature latent vector.

[0017] The detailed structure of the multi-scale dynamic convolutional residual network is as Figure 2 shown. MDCRN adopts a multi-scale convolutional neural network MSCNN architecture containing three parallel branches. Each branch has a dynamic convolutional residual module DC-ResBlock. The structure of the dynamic convolutional residual module is as Figure 3 shown. DC-ResBlock includes the main path of the residual block and the branch path of the residual connection. Dynamic convolution is used for feature extraction in the main path and the branch path respectively, and then dynamic convolution and residual connection are fused.

[0018] Step S3: Input the source domain fault feature latent vector and the target domain fault feature latent vector into the domain adaptation module, convert the extracted features of the source domain and the target domain into probability distributions, and then use the JSD loss function to calculate the JSD distribution distance value L jsd ( P , Q ).

[0019] The JSD loss function is: (1); In formula (1), P and Q are respectively the feature distributions of the source domain X s and the feature distributions of the target domain X t , M = P + Q / 2 , is Q relative to M the KL divergence, representing the difference from Q to M .

[0020] Step S4: Input the source domain fault feature latent vector into the classifier. The classifier marks some sample data as true labels and marks the remaining sample data as predicted labels through the pseudo-label method, and then uses the source domain cross-entropy loss function to calculate the error value between the true labels and the predicted labels L ( X s , Y s ), as the source domain cross-entropy loss L x .

[0021] The source domain cross-entropy loss function is: (2); In formula (2), E represents the mathematical expectation, x j s represents the true label, y j s represents the predicted label, f cross （f） represents the cross-entropy loss function.

[0022] Step S5: Input the potential vector of the target domain fault features into the classifier. The classifier marks some sample data as true labels and marks the remaining sample data as predicted labels by the pseudo-label method. Then, the error value between the true labels and the predicted labels is calculated using the target domain cross-entropy loss function L ( X t , Y t ), which is used as the target domain cross-entropy loss value L y 。

[0023] The target domain cross-entropy loss function is as follows: (3); In formula (3), E represents the mathematical expectation, x j t represents the true label, y j t represents the predicted label, f cross （f） represents the cross-entropy loss function.

[0024] Step S6: Calculate the total loss value of the sample data set using the total loss function L all ; The total loss function is as follows: (4).

[0025] Then, use the backpropagation algorithm to update the parameters of the feature extractor G f and the classifier G s . After each update, recalculate the total loss value L all , until the total loss value L all is 0 and stop the update.

[0026] The parameter update method of the feature extractor G f is as follows: (5); The classifier G s 's parameter update method is as follows: (6); In formulas (5) and (6), hDenote the learning rate, specifically set the learning rate h = 0.001, the total number of training epochs e t = 100, the batch size Batch = 256, adopt the RMSProp optimization algorithm, and use the backpropagation algorithm to update the feature extractor G f and the classifier G s parameters.

[0027] At each update in step S6, first update the cross-entropy loss value of the target domain through the weight coefficient η L y Then, calculate the total loss value through the updated L' y The calculation formula of L all , L' y is: L' y = η· L y (7); (8); In formula (8), w i = 0.001, w f = 0.1, e represents the current training epoch, e t represents the total number of training epochs.

[0028] Step S7: Through the updated feature extractor G f and the classifier G s , and cooperate with the domain adaptation module to construct the JWMS-NET joint weighted multi-scale transfer network. The feature extractor G f adopts the multi-scale dynamic convolutional residual network MDCRN to obtain rich multi-scale features. The domain adaptation module proposes an adaptive domain alignment network to align the distribution of fault features in the source domain and the target domain. After feature extraction, with the help of the classifier G s perform cross-domain fault diagnosis on the features.

[0029] Input the latent vector of the target domain fault features into the JWMS-NET neural network, and output the discriminated bearing state category of the target domain. Compare the discriminated bearing state category of the target domain with the sample data labels of the target domain to obtain the diagnostic recognition accuracy of the rolling bearing fault.

[0030] To analyze the diagnostic performance of the JWMS-NET fault diagnosis model more specifically, five models were selected for comparison, including: (a) MDA+CNN: Using a traditional convolutional neural network for feature extraction, the loss function includes: one is the source domain cross-entropy classification loss; the other is the application of the MDA distribution loss (MMD+CORAL).

[0031] (b) KL+CNN: Using a traditional convolutional neural network for feature extraction, the loss function includes: one is the source domain cross-entropy classification loss and the pseudo-label target domain classification loss (without considering adaptive weight allocation); the other is the use of the Kullback-Leibler divergence metric as the distribution loss.

[0032] (c) CORAL+CNN: Using a traditional convolutional neural network for feature extraction, the loss function includes: one is the source domain cross-entropy classification loss; the other is the application of the CORAL distribution loss.

[0033] (d) MMD+MSCNN: Using a multi-scale convolutional neural network for feature extraction, the loss function includes: one is the source domain cross-entropy classification loss; the other is the application of the MMD distribution loss.

[0034] (e) CORAL+MSCNN: Using a multi-scale convolutional neural network for feature extraction, the loss function includes: one is the source domain cross-entropy classification loss; the other is the application of the CORAL distribution loss.

[0035] MDA, MMD, KL, and CORAL are common distance metric criteria. By combining different distance metric criteria with feature extractors, cross-domain fault diagnosis is achieved. Through multiple transfer tasks, the proposed JSD distribution loss and pseudo-label cross-entropy loss mechanism are tested to verify their advantages. The network architectures and training rules of the above comparison methods are the same as those of the JWMS-NET method. Three groups of cross-machine transfer tasks are implemented to verify the accuracy and robustness of JWMS-NET diagnosis. In the following research, to avoid the contingency and randomness of diagnostic results, each experiment is conducted with 5 trials of different random initializations, and the average value of the final diagnostic results is taken based on these trials. The evaluation criteria are diagnostic accuracy, F1 score, precision, and recall.

[0036] First, dataset Ⅰ is used for verification. The target domain diagnostic results based on the WT transfer task are shown in Tables 2 and 3: Table 2 Experimental results of target domain classification accuracy based on the WT transfer task Table 3 Experimental results of F1 score, precision, and recall rate of the target domain based on 12 transfer tasks of WT As can be seen from Table 2, due to the large shift between the two domains of the variable working condition data, the performance of the basic CNN domain adaptation model is poor. Under different distribution losses of MDA, KL, and CORAL, its average accuracy only reaches 88.21%, 86.86%, and 86.35% respectively. Considering the introduction of MSCNN in the domain adaptation model, which uses multi-scale CNN for feature extraction, compared with one-dimensional concatenated CNN, the feature extraction effect is not ideal. Under different distribution losses of MMD and CORAL, its average accuracy only reaches 66.16% and 70.33% respectively. The JWMS-NET model, based on the MDCRN feature extraction, considers the source domain classification loss and the pseudo-label target domain classification loss, and makes a weighted improvement on the pseudo-label target domain classification loss. After adding the JSD distribution loss on the basis of the classification loss, the accuracy of JWMS-NET is greatly improved, and the average accuracy can reach 99.66%. As can be seen from the F1 score, precision, and recall rate results in Table 3, the diagnostic results are very stable. As can be seen from Table 3, when using JWMS-NET for diagnosis, the standard deviation difference under 12 transfer tasks of WT data is very small, and the standard deviation difference of the average recall rate is the largest, with a maximum of 1.81%. The JWMS-NET model can better handle transfer tasks under variable speeds.

[0037] Use datasets Ⅱ and Ⅲ to verify again the diagnostic reliability of the JWMS-NET model under other datasets. Its parameters are the same as the above settings. The specific transfer experiment tasks are shown in Table 1, considering 12 variable speed transfer tasks under the HUST bearing dataset and 2 variable speed transfer tasks under the Southeast University bearing fault dataset. The sample size of each fault type and the network model parameter settings are the same as those in Experiment Ⅰ. Use the method in this paper and five comparison methods to verify 14 transfer tasks. The detailed diagnostic results based on the HUST and SEU transfer tasks are shown in Tables 4, 5, and 6.

[0038] Table 4 Experimental results of the target domain classification accuracy based on the HUST transfer task Table 5 Experimental results of the target domain classification accuracy based on the SEU transfer task Table 6 Experimental results of the F1 score, precision, and recall rate of the target domain based on the HUST and SEU transfer tasks It can be seen from this the experimental results based on the HUST bearing dataset and the Southeast University bearing fault dataset. 14 transfer tasks are constructed under the two public datasets, considering the change in rotational speed in the transfer experimental tasks. The performance of the basic CNN domain adaptation model is poor. Under different distribution losses of MDA, KL, and CORAL, as shown in Table 4, the average accuracy rates on the HUST bearing dataset reach only 66.98%, 75.99%, and 71.36% respectively. On the Southeast University bearing fault dataset, the average accuracy rates reach only 78.79%, 61.04%, and 93.31% respectively. Considering the introduction of MSCNN in the domain adaptation model, which uses multi-scale CNN for feature extraction, the results are not ideal. Under different distribution losses of MMD and CORAL, on the HUST dataset, the average accuracy rates reach only 43.26% and 46.91% respectively, and on the Southeast University dataset, the average accuracy rates reach only 46.02% and 68.01% respectively. In addition, according to the evaluation metrics of F1 score, precision, and recall rate, the detailed diagnosis results on the two datasets are shown in Table 6. Among them, the data in the left column are the calculation results on the HUST dataset, and the data in the right column are the calculation results on the SEU dataset.

[0039] It can be found from the comparison of the six methods in Tables 4 and 5 that the JWMS-NET method achieves an ideal diagnosis effect with the average accuracy rates of 99.84% and 99.41% on the two datasets by introducing the pseudo-label target domain classification loss and the JSD distribution loss. Compared with the other five methods, JWMS-NET has better reliability.

[0040] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on a joint weighted multi-scale transfer algorithm, characterized in that It includes the following steps: Step S1: Obtain the vibration signals of the rolling bearing under multiple different rotational speed conditions through a public dataset, and segment the vibration signals by the data overlapping segmentation method to obtain the source domain sample dataset and the target domain sample dataset; Step S2: Input the source domain sample dataset and the target domain sample dataset into the multi-scale dynamic convolutional residual network MDCRN respectively to obtain the source domain fault feature latent vector and the target domain fault feature latent vector; Step S3: Input the source domain fault feature latent vector and the target domain fault feature latent vector into the domain adaptation module, convert the extracted features of the source domain and the target domain into probability distributions, and then use the JSD loss function to calculate the JSD distribution distance value L jsd ( P , Q ); The JSD loss function is: (1)； In formula (1), P and Q are the characteristic distributions of the source domain X s and the target domain X t , M = P + Q / 2 , is Q the KL divergence with respect to M , representing the difference from Q to M ; Step S4: Input the potential vector of the source domain fault features into the classifier. The classifier marks some sample data as true labels and marks the remaining sample data as predicted labels by the pseudo-label method. Then, the error value between the true labels and the predicted labels is calculated using the source domain cross-entropy loss function L ( X s , Y s ), which is used as the source domain cross-entropy loss L x ; The source domain cross-entropy loss function is: (2)； In formula (2), E represents the mathematical expectation, x j s represents the true label, y j s represents the predicted label, f cross （f） represents the cross-entropy loss function; Step S5: Input the potential vector of the target domain fault feature into the classifier. The classifier marks some sample data as true labels and marks the remaining sample data as predicted labels through the pseudo-label method. Then, the error value between the true label and the predicted label is calculated using the target domain cross-entropy loss function L ( X t , Y t ), which is used as the target domain cross-entropy loss value L y ; The target domain cross-entropy loss function is: (3)； In formula (3), E represents the mathematical expectation, x j t represents the true label, y j t represents the predicted label, f cross （f） represents the cross-entropy loss function; Step S6: Calculate the total loss value of the sample data set using the total loss function L all ; The total loss function is: (4)； Then, the backpropagation algorithm is used to update the parameters of the feature extractor G f and the classifier G s After each update, the total loss value is recalculated L all until the total loss value L all reaches 0 and the update stops; Feature extractor G f The parameter update method is as follows: (5)； Classifier G s The parameter update method is as follows: (6)； In formulas (5) and (6), h represents the learning rate; Step S7: Through the updated feature extractor G f and the classifier G s , and cooperate with the domain adaptation module to construct the JWMS-NET joint weighted multi-scale transfer network. Input the target domain fault feature latent vector into the JWMS-NET neural network, and output the discriminated bearing state category of the target domain. Compare and calculate the discriminated bearing state category of the target domain with the sample data label of the target domain to obtain the diagnostic recognition accuracy of the rolling bearing fault.

2. The rolling bearing fault diagnosis method based on the joint weighted multi-scale migration algorithm according to claim 1, characterized in that: In step S1, obtain the vibration signals of the rolling bearing at different rotational speeds through the WT-planetary-gearbox-dataset public dataset. The rolling bearing includes four states and operates under four working conditions. The four states include healthy rolling gear, gear damage, tooth root fracture, and missing teeth. The four working conditions include 20Hz, 30Hz, 40Hz, and 50Hz.

3. The rolling bearing fault diagnosis method based on the combined weighted multi-scale migration algorithm according to claim 1, characterized in that: In step S1, obtain the vibration signals of the rolling bearing at different rotational speeds through the bearing dataset of Huazhong University of Science and Technology. The rolling bearing includes nine states and operates under four working conditions. Select four states including normal, severe inner ring fault, severe rolling element fault, and severe compound fault. The four working conditions include 65Hz, 70Hz, 75Hz, and 80Hz.

4. The rolling bearing fault diagnosis method based on the combined weighted multi-scale migration algorithm according to claim 1, wherein: In step S1, obtain the vibration signals of the rolling bearing at different rotational speeds through the bearing fault dataset of Southeast University. The rolling bearing includes five states and operates under two working conditions. Select four states including roller fault, inner ring fault, compound fault, and normal operation. The two working conditions include 20Hz and 30Hz.

5. The rolling bearing fault diagnosis method based on the joint weighted multi-scale migration algorithm according to claim 1, characterized in that: The multi-scale dynamic convolutional residual network MDCRN in step S2 adopts the multi-scale convolutional neural network MSCNN architecture including three parallel branches. Each branch has a dynamic convolutional residual module DC-ResBlock. The dynamic convolutional residual module DC-ResBlock includes the main path of the residual block and the branch path of the residual connection. The main path and the branch path respectively use dynamic convolution for feature extraction, and then fuse the dynamic convolution and the residual connection.

6. The rolling bearing fault diagnosis method based on the combined weighted multi-scale migration algorithm according to claim 1, characterized in that: At each update in step S6, first update the cross-entropy loss value of the target domain by the weight coefficient η L y and then calculate the total loss value by the updated L' y . The calculation formula of L all is as follows: L' y ​ L' y = η· L y (7); (8)； In Equation (8), w i = 0.001, w f = 0.1, e represents the current training cycle, e t represents the total number of training cycles.

Images