Bearing fault diagnosis method and device based on dynamic domain adaptation network, and medium

Abstract

The invention relates to the field of fault diagnosis, and discloses a bearing fault diagnosis method and device based on a dynamic domain adaptation network, and a medium, and the method comprises the steps: obtaining the vibration signal data of bearings of different models under different working conditions, and taking the data as the data of a source domain and a target domain; a generative adversarial network based on time-frequency feature structure similarity is adopted to carry out data expansion, and a 1-D CNN network containing a CBAM module and a dynamic balance domain adaptation module is utilized to carry out feature extraction and classification. Source domain data are input into the network for feature extraction, and meanwhile, the cross entropy loss and the cross-domain feature are combined to adapt to the loss optimization model. And after model training is completed, target domain test data is input into the trained fault diagnosis network, a verification result shows that the model has high diagnosis accuracy and generalization ability on different data sets, and the reliability of rolling bearing fault diagnosis is remarkably improved.

Description

Technical Field

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

[0002] Bearings are widely used in almost all types of rotating machinery as key components in mechanical transmission systems. The health of bearings is directly related to their normal operation. The normal operation of rotating machinery is highly dependent on the health of bearings. Compared with other mechanical components, bearings are more prone to failure due to their complex mechanical structure and long-term operation under harsh conditions. When a rolling bearing fails seriously, it will have a serious impact on the entire operating system. Therefore, in order to reduce maintenance costs and avoid safety accidents, it is necessary to diagnose the fault in a timely and accurate manner before the fault occurs.

[0003] The core idea of ​​transfer learning is to learn common features from several related but different tasks and apply the shared knowledge to solve new problems. In actual bearing fault diagnosis applications, bearing fault data is scarce and far less than the normal data volume. This problem of unbalanced fault data will lead to reduced diagnostic accuracy of neural network models. Therefore, transfer learning can be introduced to solve the problem of insufficient bearing fault data. Transfer learning is an important research direction in the field of machine learning. It can be divided into four categories: instance-based, model-based, feature-based, and adversarial transfer learning. Among them, feature-based methods are widely used. This method maps data from the source domain and the target domain from the original feature space to a common space to reduce the distribution difference between them and realize unsupervised transfer learning. Research in the past few years has also confirmed the effectiveness of feature-based transfer learning in bearing fault diagnosis. However, existing transfer diagnosis methods only consider the marginal probability distribution or conditional probability distribution of target domain and source domain data, and only consider the importance of the two probability distributions when considering the two probability distributions at the same time, which will lead to a decrease in the diagnostic accuracy of the model in the target domain. In addition, most of the current deep learning models only consider the migration under different working conditions on the same rotating machinery, and rarely pay attention to the migration between different rotating machinery. At the same time, in actual engineering environments, it is difficult to obtain a sufficient number of well-labeled fault samples. Therefore, it is necessary to use a generative network to generate realistic fault samples to reduce the model's demand for large-scale labeled data. Summary of the invention

[0004] The purpose of the present invention is to propose a bearing fault diagnosis method, device and medium based on a dynamic domain adaptive network to solve the technical problems of insufficient samples and low accuracy of fault diagnosis results in existing fault diagnosis.

[0005] Specifically, the present invention provides a bearing fault diagnosis method based on a dynamic domain adaptive network, comprising the following steps:

[0006] S1. Collect bearing vibration signals under different models and working conditions, perform sliding window segmentation on the collected data set to form a data set containing labels, and divide the source domain and target domain data sets;

[0007] S2. Based on the similarity of time-frequency feature structures, a generative adversarial network model is built for data expansion;

[0008] S3. Build a one-dimensional convolutional neural network 1-D CNN bearing fault diagnosis framework based on convolutional block attention module CBAM;

[0009] S4. Based on the diagnosis framework in S3, the dynamic balance domain adaptation network is introduced to build the CBAM-DBDAN model. The training set data is input and the model is trained using the stochastic gradient descent algorithm. After the training is completed, the test set data is input to obtain the fault diagnosis results.

[0010] A storage medium stores instructions and data for implementing a bearing fault diagnosis method based on a dynamic domain adaptive network.

[0011] A bearing fault diagnosis device based on a dynamic domain adaptive network comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a bearing fault diagnosis method based on a dynamic domain adaptive network.

[0012] The beneficial effects provided by the present invention are: in order to solve the problems of insufficient samples and cross-domain distribution differences, a generative adversarial network based on the similarity of time-frequency feature structure is used for data expansion, and a 1-D CNN network containing a CBAM module is used for feature extraction and classification, and a dynamic factor is introduced in the domain adaptation module to measure the difference in marginal probability distribution and conditional probability distribution of source domain and target domain data, so as to reduce the distribution difference of source domain and target domain data. The migration experiment on the test set proves that the average accuracy of the method of the present invention is 98.86%, which can effectively diagnose bearing faults, play a warning role before the fault occurs, effectively avoid the occurrence of safety accidents, and reduce the cost of machine maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic flow chart of the method of the present invention;

[0014] Figure 2 This is the t-SNE result diagram of the model on the migration task B2→A0;

[0015] Figure 3 This is the t-SNE result diagram of the ablation experiment model 1 on task B2→A0;

[0016] Figure 4 It is a schematic diagram of the working of the hardware device of an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0018] Before formally describing the present invention, the scheme of the present invention is first generally described for easy understanding.

[0019] Please refer to Figure 1 The present invention provides a bearing fault diagnosis method based on a dynamic domain adaptive network, comprising the following steps:

[0020] S1. Collect bearing vibration signals under different models and working conditions, perform sliding window segmentation on the collected data set to form a data set containing labels, and divide the source domain and target domain data sets;

[0021] As an embodiment, step S1 of the present invention is specifically as follows:

[0022] S101. Collect time domain vibration signals of rolling bearings of two different models A and B, which are recorded as Vat and Vbt.

[0023] S102, using the sliding window method to segment the rolling bearing data sets Vat and Vbt of two different models A and B, the window size is window_size, and the segmented rolling bearing data of the two models are set as DATA_A and DATA_B, including labels Label{0, 1, 2, 3};

[0024] S2. Based on the similarity of time-frequency feature structures, a generative adversarial network model is built for data expansion;

[0025] It should be noted that step S2 of the present invention is specifically as follows:

[0026] S201. Build a generative adversarial network model, and divide the DATA_A data into a training set and a test set in a ratio of 8:2, and input them into the built generative adversarial network model for training and testing.

[0027] S202. The generative adversarial network consists of two parts: a generator and a discriminator. The goal of the generator is to generate new samples by learning the distribution of real data, and the discriminator is used to determine whether the input data comes from real data or is generated by the generator. The two are jointly optimized through game theory, so that the generator can generate more and more real data.

[0028] S203. In the test phase, the time-frequency characteristics of the generated data are optimized by adversarial loss and time-frequency structure loss, so that the generated data has time-frequency characteristics similar to the real data. The generator loss function is expressed as:

[0029]

[0030] In the formula, is the loss function of the generator; Express expectation; c is the input noise of the generator, ~ indicates that it obeys a specific distribution; p(z c ) is the noise distribution; D c () is the discriminator function; G c () is the generator function; θ c is the parameter of the generator; φ c is the parameter of the discriminator; L TF (G c (z c θ c )) is the time-frequency structure loss; λ c is the weight for controlling the loss of time-frequency structure; L div (G c ) is the diversity enhancement loss.

[0031] The loss function of the discriminator includes the traditional adversarial loss and time-frequency structure loss, and the calculation method is expressed as:

[0032]

[0033] In the formula, p real (x c ) is the real data distribution; φ c is the parameter of the discriminator; c is the weight of the time-frequency structure loss.

[0034] S204, the generator and the discriminator are alternately optimized by adversarial training. In a training cycle, the generator tries to generate more and more realistic samples, and the discriminator tries to improve its ability to identify true and false samples. The generator and the discriminator perform back propagation according to their respective loss functions and update their respective network weights through gradients. The update method is expressed as:

[0035]

[0036] In the formula, is the learning rate of the generative adversarial network at the tth iteration, indicating the step size of parameter update; ← is the parameter update operation; is the gradient of the generator’s loss function with respect to its parameters; is the gradient of the discriminator's loss function with respect to its parameters.

[0037] The learning rate of the generative adversarial network is updated using an adaptive gradient adjustment method, making the training of the generator and discriminator more efficient and stable, expressed as:

[0038]

[0039] In the formula, is the sum of squared gradients; ∈ cas is a small regularization constant to prevent division by zero errors; η cas is the initial learning rate of the generative adversarial network.

[0040] S205. By adopting the diversity constraint term in the loss function of the generator, the generator is required to keep the generated samples consistent with the real samples and cover different areas in the data space as much as possible, thereby increasing the diversity of the generated samples. The calculation method of the diversity enhancement loss is expressed as:

[0041]

[0042] In the formula, and are the i-th and j-th noise samples input to the generator; / / / is the L2 norm.

[0043] S206, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0044] S3. Build a one-dimensional convolutional neural network 1-D CNN bearing fault diagnosis framework based on convolutional block attention module CBAM;

[0045] It should be noted that step S3 of the present invention is specifically as follows:

[0046] S301. During the training of the network model, batch normalization is used for each small batch to solve the problem that the update of the training parameters of the previous layer causes the distribution of input data of the subsequent layer to change, thereby accelerating the convergence speed and improving the model accuracy.

[0047] S302, using maximum pooling to reduce the size of spatial information and improve computing efficiency:

[0048] in is the i-th element of the pooling region.

[0049] S303, use the fully connected layer to map the learned distributed feature representation to the sample label space, described as x l =f(w l x l-1 +b l ), where w land b l They represent the weight and bias of the fully connected layer respectively, and the ReLU unit is used as the activation function: ReLU(x)=max(0,x).

[0050] S304: Input the features obtained in step S303 into the CBAM module to assign values ​​to each channel and spatial coordinate. CBAM consists of two parts: channel attention and spatial attention. The weight calculation process of CAM and SAM is as follows:

[0051]

[0052] Among them, W CAM is the channel attention weight, W SAM represents the spatial attention weight, MLP(·) is the shared fully connected layer, avg(·) is the average value, and max(·) is the maximum value.

[0053] S4. Based on the diagnosis framework in S3, the dynamic balance domain adaptation network is introduced to build the CBAM-DBDAN model. The training set data is input and the model is trained using the stochastic gradient descent algorithm. After the training is completed, the test set data is input to obtain the fault diagnosis results.

[0054] It should be noted that step S4 of the present invention is specifically as follows:

[0055] S401. Introduce a balancing factor in the domain adaptation network to dynamically adjust the weights of the marginal probability distribution and the conditional probability distribution to reduce the distribution difference between the source domain and the target domain, and use the distance between the two distributions as the loss function in the form of summation to dynamically measure the difference between the two distributions, so that the feature distribution obtained from the target domain is similar to the source domain data distribution. Using cross entropy The extracted source domain features calculate the cross entropy loss, where Θ is the parameter of the neural network, for The corresponding output label.

[0056] S402. Calculate domain adaptation loss using cross-domain features: in Represents the output features of the fifth convolutional layer and the fully connected layer of the network, MMD m is the maximum mean difference, which is used to match the marginal probability distribution. CORAL represents the second-order statistical difference between the source domain and target domain features, which is used to match the conditional probability distribution. α∈[0,1] is the dynamic weight of the marginal probability distribution distance.

[0057] S403, these losses are combined to calculate the total loss, and the network parameters are updated by back propagation using a parameter optimization method based on stochastic gradient descent. The total loss is as follows:

[0058]

[0059] S40. Input the obtained test set data into the CBAM-DBDAN diagnostic model to obtain the diagnostic classification results of the test bearing, and evaluate the accuracy and reliability of the model based on the real bearing fault labels, and then verify the effectiveness of each module through ablation experiments.

[0060] Embodiment 1:

[0061] The present invention is further described below in conjunction with specific embodiments, which are specific embodiments of the rolling bearing fault diagnosis method of the present invention:

[0062] The data for this case comes from the bearing vibration data in the public data sets of Case Western Reserve University and Jiangnan University. The sampling frequency of the bearing data of Case Western Reserve University is 48kHz, and the sampling frequency of the bearing data of Jiangnan University is 50kHz. The bearing models include SKF6205, N205, and NU205; the motor speed is 1460RPM, 1560RPM, and 1660RPM, and the health status of the bearing includes: normal (N), inner ring fault (IF), outer ring fault (OF), and rolling element fault (BF).

[0063] (1) Collect vibration signals of bearings of different models and working conditions, and segment the bearing data sets of different models and working conditions. The sliding window method is used to segment the rolling bearing data sets Vat and Vbt of different models. The window size window_size is 3072. The segmented rolling bearing data of the two models contain labels Label{0, 1, 2, 3}. The segmented data sets are shown in Table 1:

[0064] Table 1 The data set after segmentation in this embodiment

[0065]

[0066] (2) Build a generative adversarial network model that considers the similarity of time-frequency feature structures and expand the data.

[0067] The weights and biases of the generator and the discriminator are initialized randomly, and the initialized parameters obey the normal distribution with a mean of 0 and a variance of the unit matrix. The generator and the discriminator optimize the time-frequency characteristics of the generated data through adversarial loss and time-frequency structure loss, and the time-frequency transformation function uses short-time Fourier transform to calculate the time-frequency characteristics of the signal. The generator and the discriminator perform back propagation according to their respective loss functions, update their respective network weights through gradients, and update the learning rate using an adaptive gradient adjustment method. After the sample expansion model training is completed, the trained data expansion model is used to increase the number of samples. In one embodiment, assuming that the original sample is 800, the data expansion model expands and generates 200 samples, and the expanded data set contains 1000 samples.

[0068] (3) Build a 1-D CNN bearing fault diagnosis framework based on CBAM.

[0069] Batch normalization is used for each small batch to solve the problem that the update of the training parameters of the previous layer causes the distribution of the input data of the later layer to change, which speeds up the convergence speed and improves the model accuracy. The maximum pooling is used to reduce the size of the spatial information and improve the computational efficiency. The fully connected layer is used to map the learned distributed feature representation to the sample label space, and the ReLU unit is used as the activation function. The fault diagnosis network parameters available for the target rolling bearing are shown in Table 2:

[0070] Table 2 Available fault diagnosis network parameters for target rolling bearings

[0071]

[0072] (4) The training set data is input into the CBAM-DBDAN model, and the model is trained using the stochastic gradient descent algorithm. After the training is completed, the test bearing data is input to obtain the fault diagnosis results.

[0073] Model training: The balance factor is used to dynamically adjust the weights of the marginal probability distribution and the conditional probability distribution. When the data distribution difference is large, the marginal probability distribution is considered first. When the data distribution difference is small, the conditional probability distribution is given priority. The distance between the two distributions is taken as the loss function in the form of summation, so as to dynamically measure the distribution difference so that the target domain feature distribution is similar to the source domain. Furthermore, the cross entropy is used to extract the source domain features to calculate the cross entropy loss, and the domain adaptation loss is calculated based on the cross-domain features. The dynamic weights are adjusted according to the output features of the network convolutional layer and the fully connected layer. Finally, the total loss is calculated by combining these losses, and the network parameters are updated based on the stochastic gradient descent algorithm. The 1-D CNN weights and biases are dynamically adjusted. After multiple iterations, the model training is completed until the preset stop condition is reached.

[0074] Model testing: The migration scenario of this case takes into account the following two types of differences: different bearings and different working conditions. The target domain data is input into the constructed bearing fault diagnosis model to obtain the target domain bearing fault diagnosis results. The fault diagnosis accuracy of the target domain bearing is shown in Table 3, and the average diagnostic accuracy is 98.86%. In order to better illustrate the effectiveness of the proposed method, taking the B2→A0 migration task as an example, the features extracted by the model are used as output, and the high-dimensional data is reduced and visualized using T-SNE. It can be seen that various types of data in the source domain and the target domain can be effectively aligned, the edge interval is large, and the classification boundary is clear, which can effectively diagnose the bearing fault type of the motor. The results are as follows: Figure 2 shown.

[0075] Table 3. Bearing fault diagnosis accuracy in target domain

[0076] A0→B1 B1→A0 A0→B2 B2→A0 Average accuracy Accuracy (%) 98.48 98.76 99.35 98.88 98.86

[0077] At the same time, in order to illustrate the effectiveness of each module of the present invention, an ablation experiment was conducted in this case, and the model setting is described as follows:

[0078] Model 1: Compared with the model proposed in this invention, the CBAM attention mechanism is not added.

[0079] Model 2: Same feature extraction network structure as the model proposed in this invention, but without dynamic domain adaptation.

[0080] Model 3: has the same structure as the model proposed in this invention, but does not use the generative adversarial network to supplement sample data.

[0081] The fault diagnosis results of each ablation model are shown in Table 4. It can be seen that the diagnosis accuracy of the model proposed in this scheme is much higher than that of the ablation model.

[0082] Table 4 Fault diagnosis accuracy of each ablation model

[0083]

[0084] Taking the migration task B2→A0 as an example, the t-SNE result of model 1 in the ablation experiment is as follows: Figure 3 As shown in Figure 2, it can be clearly seen that the various types of data in the source domain and the target domain cannot be effectively aligned, the edge interval is small, and the classification boundary is fuzzy.

[0085] Embodiment 2:

[0086] See also Figure 4 , Figure 4 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a bearing fault diagnosis device 401 based on a dynamic domain adaptive network, a processor 402 and a storage medium 403.

[0087] A bearing fault diagnosis device 401 based on a dynamic domain adaptive network: The bearing fault diagnosis device 401 based on a dynamic domain adaptive network implements the bearing fault diagnosis method based on a dynamic domain adaptive network.

[0088] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the bearing fault diagnosis method based on the dynamic domain adaptive network.

[0089] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the bearing fault diagnosis method based on the dynamic domain adaptive network.

[0090] The beneficial effects of the present invention are as follows: in order to solve the problems of insufficient samples and cross-domain distribution differences, a generative adversarial network based on the similarity of time-frequency feature structure is used for data expansion, and a 1-DCNN network containing a CBAM module is used for feature extraction and classification, and a dynamic factor is introduced into the domain adaptation module to measure the difference in marginal probability distribution and conditional probability distribution of source domain and target domain data, so as to reduce the distribution difference of source domain and target domain data. The migration experiment on the test set proves that the average accuracy of the method of the present invention is 98.86%, which can effectively diagnose bearing faults, play a warning role before the fault occurs, effectively avoid the occurrence of safety accidents, and reduce the cost of machine maintenance.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method based on a dynamic domain adaptive network, characterized in that: The following steps are involved: S1. Collect bearing vibration signals under different models and working conditions, perform sliding window segmentation on the collected data set to form a data set containing labels, and divide the source domain and target domain data sets; S2. Based on the similarity of time-frequency feature structures, a generative adversarial network model is built for data expansion; S3. Build a one-dimensional convolutional neural network 1-D CNN bearing fault diagnosis framework based on convolutional block attention module CBAM; S4. Based on the diagnosis framework in S3, the dynamic balance domain adaptation network is introduced to build the CBAM-DBDAN model. The training set data is input and the model is trained using the stochastic gradient descent algorithm. After the training is completed, the test set data is input to obtain the fault diagnosis results.

2. A bearing fault diagnosis method based on a dynamic domain adaptive network as claimed in claim 1, characterized in that: Step S1 is specifically as follows: The collected bearing vibration signal is segmented by the sliding window method, where the window size is window_size. The segmented data sets are labeled DATA_A and DATA_B, respectively, and contain fault labels Label{0, 1, 2, 3}.

3. A bearing fault diagnosis method based on a dynamic domain adaptive network as claimed in claim 2, characterized in that: S21, build a generative adversarial model, and divide the data set DATA_A into a training set and a test set according to the proportion, and input them into the built generative adversarial model for training and testing; S22. In the training phase, the generative adversarial model consists of two parts: a generator and a discriminator. The goal of the generator is to generate new samples by learning the distribution of real data, and the discriminator is used to determine whether the input data comes from real data or is generated by the generator. The two are jointly optimized through game theory, so that the generator can generate more and more real data. S23. During the testing phase, in order to ensure that the generated data can fully reflect the time domain and frequency domain information of the real data, the time-frequency structure loss is added to the loss function of the generative adversarial network model taking into account the similarity of the time-frequency feature structure.

4. A bearing fault diagnosis method based on a dynamic domain adaptive network as claimed in claim 3, characterized in that: The generator loss function is expressed as: In the formula, is the loss function of the generator; Express expectation; c is the input noise of the generator, ~ indicates that it obeys a specific distribution; p(z c ) is the noise distribution; D c () is the discriminator function; G c () is the generator function; θ c is the parameter of the generator; φ c is the parameter of the discriminator; L TF (G c (z c θ c )) is the time-frequency structure loss; λ c is the weight for controlling the loss of time-frequency structure; L div (G c ) is the diversity enhancement loss; The loss function of the discriminator includes the traditional adversarial loss and time-frequency structure loss, and the calculation method is expressed as: In the formula, p real (x c ) is the real data distribution; φ c is the parameter of the discriminator; c is the weight of the time-frequency structure loss.

5. The bearing fault diagnosis method based on dynamic domain adaptive network according to claim 1, characterized in that: Step S3 is as follows: S31, performing batch normalization processing on the input data; S32, reducing the size of spatial information through a maximum pooling operation; S33, use the fully connected layer to map the features to the fault label space, and use the ReLU activation function to optimize the network performance; S34. Improve the feature extraction capability of the model by integrating channel attention and spatial attention mechanisms.

6. A bearing fault diagnosis method based on a dynamic domain adaptive network as claimed in claim 5, characterized in that: Step S4 is specifically as follows: S41, calculating the cross entropy loss using the source domain features extracted by the cross entropy; S42, introduce a balancing factor to dynamically adjust the weights of the marginal probability distribution and the conditional probability distribution to reduce the distribution difference between the source domain and the target domain; S43, combining the cross entropy loss and the cross-domain feature adaptation loss to calculate the total loss, and using the parameter optimization method based on stochastic gradient descent to update the network parameters through back propagation; S44. Input the obtained test set data into the CBAM-DBDAN diagnostic model to obtain the diagnostic classification results of the test bearing, and evaluate the accuracy and reliability of the model based on the real bearing fault labels, and then verify the effectiveness of each module through ablation experiments.

7. A bearing fault diagnosis method based on a dynamic domain adaptive network as claimed in claim 6, characterized in that: The total loss in step S43 is as follows: in, is the cross entropy loss, Θ is the parameter of the neural network, for The corresponding output label: in Represents the output features of the fifth convolutional layer and the fully connected layer of the network, MMD m is the maximum mean difference, which is used to match the marginal probability distribution. CORAL represents the second-order statistical difference between the source domain and target domain features, which is used to match the conditional probability distribution. α∈[0,1] is the dynamic weight of the marginal probability distribution distance.

8. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a bearing fault diagnosis method based on a dynamic domain adaptive network as described in any one of claims 1 to 7.

9. A bearing fault diagnosis device based on a dynamic domain adaptive network, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement a bearing fault diagnosis method based on a dynamic domain adaptive network as described in any one of claims 1 to 7.

Images