Small-sample cross-domain bearing fault diagnosis method and system based on improved model independent element learning
Through improved model-independent meta-learning method and parallel synergistic attention feature encoder, the problems of insufficient samples and variability in bearing fault diagnosis are solved, and the accuracy and robustness of bearing fault diagnosis are improved, thus achieving the following: Efficient fault identification under small sample conditions.
Patent Information
- Application Number
- CN202410917134.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-10
- Publication Date
- 2024-11-08
AI Technical Summary
In actual industrial environments, the accuracy of bearing fault diagnosis is low, mainly due to the inability to obtain sufficient fault signals and the complex and variability of mechanical equipment working conditions, resulting in poor diagnosis effect of deep learning methods in the case of insufficient samples.
Using an improved model-independent element learning method, the self-similarity of the original vibration signal of the bearing is converted into a two-dimensional recursive graph, and a feature encoder based on parallel coordinated attention is constructed, using support sets and query sets for inner and outer rings. Training, add training regulatory variables to optimize the model and improve diagnostic accuracy.
The accuracy of bearing fault diagnosis under the cross-domain conditions of few samples is improved, feature extraction is achieved that is insensitive to operating conditions changes is enhanced, diagnostic capabilities in actual industrial scenarios are enhanced, and the average diagnostic accuracy rate of 97.76%.
Smart Images

Figure CN118913694A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of mechanical fault diagnosis, and in particular to a cross-domain bearing fault diagnosis method under sample-limited conditions based on improved model-independent meta-learning. Background Art
[0002] In the modern industrial field, bearings, as important components of rotating machinery, have played a vital role in high-speed trains, aerospace, electricity and other fields. Therefore, fault diagnosis methods have to cope with more demanding diagnostic scenarios: (1) Insufficient fault data: For critical mechanical equipment, their safety and stability are relatively high, and they are usually in a healthy state for a long time, so it is difficult to collect fault status data, and it is necessary to complete the diagnostic task with a limited number of samples. (2) Variability of working conditions: Production tasks are diverse, and data needs to be collected from different machines or the same mechanical equipment under different working conditions (such as speed, load, etc.) to complete specific types of diagnostic tasks.
[0003] In recent years, with the rapid development of deep learning technology, deep learning models have also been applied to the field of intelligent fault diagnosis to achieve convenient, efficient and accurate monitoring and diagnosis, instead of relying on expert experience and manual feature extraction operations. However, the success of most deep learning methods depends on a large amount of labeled training data that is the same distribution as the test data. Obviously, in actual engineering scenarios, mechanical equipment that has been in a healthy working state for a long time cannot meet this data demand. Transfer learning can effectively reduce the difference in data distribution between different working conditions, learn knowledge from one task, and then transfer the learned knowledge to different but similar tasks, providing a way to solve the above-mentioned conditional restrictions. However, when there are very few samples, the transfer learning algorithm is prone to negative transfer, and the network before the transfer is trained for easy generalization. Transfer learning only works "accidentally" and may not be optimal. Summary of the invention
[0004] The purpose of the present invention is to solve the problem of low fault diagnosis accuracy caused by the inability to obtain sufficient fault signals in actual industrial environments and the complex and changeable working conditions of mechanical equipment, and proposes a few-sample cross-domain bearing fault diagnosis method and system based on improved model-independent meta-learning.
[0005] The specific process of the few-sample cross-domain bearing fault diagnosis method based on improved model-independent meta-learning is as follows:
[0006] A small number of original vibration signals of bearings to be diagnosed are obtained, the self-similarity of the original vibration signals of the bearings is calculated to convert them into a two-dimensional recursive graph, and the bearing fault diagnosis method based on improved model-independent meta-learning is applied to the two-dimensional recursive graph to obtain the diagnosis results of the bearing fault type;
[0007] The bearing fault diagnosis method is obtained by:
[0008] Step 1: Obtain the original vibration signal of the bearing, calculate the self-similarity of the original vibration signal of the bearing, convert it into a two-dimensional recursive graph, and divide it into a support set and a query set;
[0009] Step 2: Improve the model-independent meta-learning method and add training control variables to the inner-loop training process;
[0010] Step 3: Construct a feature encoder based on parallel collaborative attention;
[0011] The feature encoder is composed of the following modules connected in sequence: a two-dimensional convolution module, a parallel collaborative attention module, and a fully connected layer;
[0012] The two-dimensional convolution module is used to extract local fault features of the two-dimensional recursive graph of the vibration signal;
[0013] The parallel collaborative attention module is used to simultaneously extract common diagnostic knowledge among multiple dimensions of length, width, and channel;
[0014] The fully connected layer is used to classify the fault features output by the parallel collaborative attention module;
[0015] Step 4: Use the support set and the query set to train the parallel collaborative attention-based feature encoder constructed in step 3 in the inner loop and outer loop of meta-learning, respectively, to obtain a trained parallel collaborative attention-based feature encoder;
[0016] Step 5: Use cross-domain vibration signals under other working conditions to test the trained parallel collaborative attention-based feature encoder obtained in step 4 to obtain a few-sample cross-domain bearing fault diagnosis method.
[0017] Furthermore, the step 1 includes the following steps:
[0018] Firstly, the sample extraction parameters are set to collect the original vibration signal of the bearing and obtain a small number of samples;
[0019] Then, the self-similarity of the calculated signal is converted into a two-dimensional recursive graph, and a small number of processed samples are randomly selected to be divided into a support set and a query set, and the category labels are set respectively.
[0020] Furthermore, the parameter optimization process of the improved model-independent meta-learning method in step 2 is:
[0021]
[0022] where θ is the initial parameter of the feature encoder based on parallel collaborative attention, θ i″ is the parameter updated after inner loop training, δ is the training control variable, α and β are the inner loop and outer loop learning rates respectively, L Ti It's a mission loss.
[0023] Furthermore, the output of the two-dimensional convolution module in step 3 is as follows:
[0024] F=Maxpooling(BN(ReLU(Conv(x)))) (4)
[0025] Among them, F is the output of the two-dimensional convolution module, x is the input of the two-dimensional convolution module, Maxpooling() is the maximum pooling operation, BN() is batch normalization, ReLU is the activation function, and Conv() is the convolution operation.
[0026] Furthermore, the output of the parallel collaborative attention module in step 3 is as follows:
[0027]
[0028] in, is the feature map that aggregates the other two spatial dimensions, are two channel feature statistics, μ and ω are trainable floating parameters, is the attention weight, F′ W , F′ H 、F C ′ is the enhanced feature map in three dimensions: width, height, and channel, and F′ is the final feature map output by the parallel collaborative attention module.
[0029] Furthermore, the method updates the network parameters through the Adam optimizer, specifically:
[0030]
[0031] Among them, θ t is the network parameter at time t, θ t+1 is the network parameter at time t+1, α is the learning rate, is the first-order moment estimate of the gradient, is the second moment estimate of the gradient, and ε is a small positive value.
[0032] A small number of cross-domain bearing fault diagnosis system based on improved model-independent meta-learning, characterized in that: the system includes: a data acquisition and extraction module, a data processing and division module, a feature encoder building module, an improved model-independent meta-learning training module, a testing module, and a fault state recognition module;
[0033] The data acquisition and extraction module is used to extract a small amount of samples, namely the original vibration signal of the bearing;
[0034] The data processing and partitioning module is used to calculate the self-similarity of the original vibration signal and convert it into a two-dimensional recursive graph, which is divided into a support set and a query set;
[0035] The feature encoder building module is used to build a feature encoder based on parallel collaborative attention;
[0036] The improved model-independent meta-learning training module is used to train a feature encoder based on parallel collaborative attention using a support set and a query set to obtain a trained feature encoder;
[0037] The testing module is used to test the trained parallel collaborative attention-based feature encoder using cross-domain vibration signals under other working conditions;
[0038] The fault status identification module is used to identify the original vibration signal of the bearing and diagnose the bearing fault type.
[0039] The beneficial effects of the present invention are:
[0040] The present invention proposes to calculate the self-similarity of the signal to process the original vibration signal, and obtains training data that is rich in features and insensitive to changes in working conditions. The present invention improves the model-independent meta-learning method and adds training control variables to the inner loop training process to optimize the training process. The present invention constructs a feature encoder based on parallel collaborative attention, and simultaneously models attention in multiple dimensions of length, width, and channel to effectively extract diagnostic knowledge. Improving the fault diagnosis accuracy under cross-domain conditions with few samples is more in line with actual industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a flow chart of the present invention;
[0042] Figure 2 is the two-dimensional recurrence diagram of the vibration signal;
[0043] Figure 3 for an improved model-independent meta-learning parameter optimization process;
[0044] Figure 4 It is a parallel collaborative attention module;
[0045] Figure 5 This is a picture of the experimental bench;
[0046] Figure 6 Pictures of rolling bearings used in the experiment;
[0047] Figure 7 is the confusion matrix of the diagnosis result of the present invention;
[0048] Figure 8 A two-dimensional cluster diagram of the diagnosis results of the present invention;
[0049] Fig. 9 This is the comparison result between the present invention and other methods. DETAILED DESCRIPTION
[0050] Specific implementation method 1: Figure 1 As shown, this implementation method adopts a method different from the traditional deep learning-based method to diagnose bearing faults based on the characteristics of bearing fault diagnosis. The specific process of the proposed small-sample cross-domain bearing fault diagnosis method based on improved model-independent meta-learning is as follows: obtaining a two-dimensional recursive graph of the original vibration signal of the bearing to be diagnosed, inputting the two-dimensional recursive graph of the original vibration signal of the bearing to be diagnosed into a parallel collaborative network application improved model-independent meta-learning method, and obtaining the bearing fault category diagnosis result.
[0051] The few-sample cross-domain fault diagnosis method is obtained by:
[0052] Step 1: Obtain the original vibration signal of the bearing and convert it into a two-dimensional recursive graph, which is divided into a support set and a query set:
[0053] Firstly, the sample extraction parameters are set to collect the original vibration signal of the bearing and obtain a small number of samples;
[0054] Then, if Figure 1 As shown, the self-similarity of the calculated signal is converted into a two-dimensional recursive graph, and a small number of processed samples are randomly selected to be divided into a support set and a query set, and the category labels are set respectively.
[0055] The samples have richer fault characteristics and are not affected by changes in operating conditions.
[0056] Step 2: Improve the model-independent meta-learning method and add training control variables to the inner-loop training process;
[0057] The parameter optimization process of the improved model-independent meta-learning method is as follows: Figure 2 As shown, specifically:
[0058]
[0059] where θ is the initial parameter of the feature encoder based on parallel collaborative attention, θ i ″ is the parameter updated after inner loop training, δ is the training control variable, α and β are the inner loop and outer loop learning rates respectively, It's a mission loss.
[0060] Step 3: Construct a feature encoder based on parallel collaborative attention;
[0061] The feature encoder is composed of the following modules connected in sequence: a two-dimensional convolution module, a parallel collaborative attention module, and a fully connected layer;
[0062] The two-dimensional convolution module includes: a convolution layer, a batch normalization layer, a maximum pooling layer, and a ReLU activation function, which are used to extract local fault features of the two-dimensional recursive graph of the vibration signal. The output of the two-dimensional convolution module is as follows:
[0063] F=Maxpooling(BN(ReLU(Conv(x)))) (12)
[0064] Among them, F is the output of the two-dimensional convolution module, x is the input of the two-dimensional convolution module, Maxpooling() is the maximum pooling operation, BN() is batch normalization, ReLU is the activation function, and Conv() is the convolution operation.
[0065] like Figure 3 As shown, the input of the parallel collaborative attention module is the output of the two-dimensional convolution module, which is used to simultaneously extract common diagnostic knowledge between multiple dimensions of length, width, and channel. The specific output is:
[0066]
[0067] in, is the feature map that aggregates the other two spatial dimensions, are two channel feature statistics, μ and ω are trainable floating parameters, is the attention weight, F′ W , F′ H 、F C ′ is the enhanced feature map in three dimensions: width, height, and channel, and F′ is the final feature map output by the parallel collaborative attention module.
[0068] Use the Adam optimizer to update the network parameters, specifically:
[0069]
[0070] Among them, θ t is the network parameter at time t, θ t+1 is the network parameter at time t+1, α is the learning rate, is the first-order moment estimate of the gradient, is the second moment estimate of the gradient, and ε is a small positive value.
[0071] In this step, the parallel collaborative attention mechanism is embedded in the convolutional neural network to build a feature encoder based on parallel collaborative attention, which can simultaneously capture the interactive relationship between the three dimensional features of channel, height and width, and effectively extract the common diagnostic knowledge of multiple tasks.
[0072] Step 4: Use the support set and query set to train the feature encoder based on parallel collaborative attention in the inner loop and outer loop of meta-learning respectively to obtain the trained feature encoder;
[0073] Step 5: Use cross-domain vibration signals under other working conditions to test the trained parallel collaborative attention-based feature encoder to obtain a few-sample cross-domain bearing fault diagnosis method;
[0074] A few-sample cross-domain bearing fault diagnosis method based on improved model-independent meta-learning can realize intelligent diagnosis and classification of bearing faults under few-sample cross-domain conditions. The collected fault signals are processed and then applied to obtain real-time diagnosis results. At the same time, this algorithm can be applied to online monitoring and diagnosis systems under the condition of equipment hardware support to perform real-time status monitoring and fault diagnosis of bearings.
[0075] Specific implementation method 2: A small number of sample cross-domain bearing fault diagnosis system based on improved model-independent meta-learning, comprising: a data acquisition and extraction module, a data processing and division module, a feature encoder building module, an improved model-independent meta-learning training module, a testing module, and a fault state recognition module;
[0076] The data acquisition and extraction module is used to set certain sample extraction parameters to extract a small amount of samples, namely the original vibration signal of the bearing;
[0077] The data processing and partitioning module is used to calculate the self-similarity of the original vibration signal and convert it into a two-dimensional recursive graph, which is divided into a support set and a query set;
[0078] The feature encoder building module is used to build a feature encoder based on parallel collaborative attention;
[0079] The improved model-independent meta-learning training module is used to train a feature encoder based on parallel collaborative attention using a support set and a query set to obtain a trained feature encoder;
[0080] The testing module is used to test the trained parallel collaborative attention-based feature encoder using cross-domain vibration signals under other working conditions;
[0081] The fault status identification module is used to identify the original vibration signal of the bearing and diagnose the bearing fault type.
[0082] Example:
[0083] The operating status of mechanical equipment changes with the change of production tasks. In order to verify the effectiveness of the proposed method under the condition of few samples and cross-domain, the following rolling bearing fault identification experiment is designed:
[0084] 1. Experimental Data
[0085] Bearing faults are collected from the fault diagnosis test bench, such as Figure 4 As shown. The sampling frequency is 12kHz, and three working conditions are set, namely C1 (speed is 1797RPM, motor load is 0HP), C2 (speed is 1772RPM, motor load is 1HP) and C3 (speed is 1750RPM, motor load is 2HP). Each working condition includes 4 fault types, among which there are 3 fault sizes except for the normal healthy state bearing, and each health state is marked with 0 to 9. Figure 5 Normal and faulty bearings are shown and detailed information is given in Table 1.
[0086] Table 1 Rolling bearing specific fault description table
[0087]
[0088] 2. Experimental results and analysis
[0089] As shown in Table 2, under six cross-operating scenarios and different fault category-sample quantity settings, the average diagnostic accuracy of the method proposed in the present invention reached 97.76%, 95.52%, 99.00%, and 95.50%, respectively, indicating that the present invention can achieve good diagnostic effects in different cross-domain scenarios with only a small number of samples, proving the effectiveness and robustness of the proposed method.
[0090] Table 2 Diagnostic accuracy of the present invention
[0091]
[0092] In order to clearly demonstrate the effectiveness of the method proposed in the present invention, confusion matrix and two-dimensional clustering diagram are used to visualize the bearing fault diagnosis results of the method proposed in the present invention. Figure 6 As shown in Figure 1, except for health states 4 and 7, the fault diagnosis accuracy of bearings in other health states reaches 100%. Figure 7 As shown in the figure, there are only 10 samples in each health state, but the rolling bearings in each health state are relatively concentrated, and obvious boundaries can be seen. Overall, the method proposed in this paper has a good fault diagnosis effect under the conditions of few samples and cross-domain.
[0093] In order to further verify the superiority of the method proposed in this invention, the method is compared with other related deep learning methods and meta-learning methods. Figure 8 As shown, whether compared with deep learning methods or meta-learning methods, the diagnostic performance of the present invention in various cross-domain scenarios is better than other methods.
[0094] In summary, the present invention provides a few-sample cross-domain bearing fault diagnosis method and system based on improved model-independent meta-learning. First, the self-similarity of the original vibration signal is calculated to convert it into a two-dimensional recursive graph that is feature-rich and not affected by changes in working conditions. A feature encoder based on parallel collaborative attention is constructed to simultaneously capture the interactive relationship between the three-dimensional features of channel, height and width, and effectively extract common diagnostic knowledge for multiple tasks. Finally, the training control variables are added to the inner-loop training process to improve the model-independent meta-learning method and improve the accuracy of few-sample cross-domain rolling bearing fault diagnosis.
Claims
1. A few-sample cross-domain bearing fault diagnosis method based on improved model-independent meta-learning, Features The specific process of the described method is: obtaining the original vibration signal of the bearing to be diagnosed, converting the original vibration signal of the bearing to be diagnosed into a two-dimensional recursive graph by autocorrelation calculation, and applying the two-dimensional recursive graph to the bearing fault diagnosis method to obtain the diagnosis result of the bearing fault type; The bearing fault diagnosis network is obtained by: Step 1: Obtain the original vibration signal of the bearing, and convert the original vibration signal of the bearing into a two-dimensional recursive graph, which is divided into a support set and a query set; Step 2: Improve the model-independent meta-learning method and add training control variables to the inner-loop training process; Step 3: Construct a feature encoder based on parallel collaborative attention; The feature encoder is composed of the following modules connected in sequence: a two-dimensional convolution module, a parallel collaborative attention module, and a fully connected layer; The two-dimensional convolution module is used to extract local fault features of the two-dimensional recursive graph of the vibration signal; The parallel collaborative attention module is used to simultaneously extract common diagnostic knowledge among multiple dimensions of length, width, and channel; The fully connected layer is used to classify the fault features output by the parallel collaborative attention module; Step 4: Use the support set and the query set to train the parallel collaborative attention-based feature encoder constructed in step 3 in the inner loop and outer loop of meta-learning, respectively, to obtain a trained parallel collaborative attention-based feature encoder; Step 5. Use cross-domain data under other working conditions to test the trained parallel collaborative attention-based feature encoder obtained in step 4 to obtain a few-sample cross-domain bearing fault diagnosis method.
2. According to claim 1, the improved model-independent meta-learning based few-sample cross-domain bearing fault diagnosis method, Features: The step 1 comprises the following steps: Firstly, the sample extraction parameters are set to collect the original vibration signal of the bearing and obtain a small number of samples; Then, the self-similarity of the calculated signal is converted into a two-dimensional recursive graph, and a small number of processed samples are randomly selected to be divided into a support set and a query set, and the category labels are set respectively.
3. According to claim 2, the improved model-independent meta-learning based few-sample cross-domain bearing fault diagnosis method, Features: The parameter optimization process of the improved model-independent meta-learning method in step 2 is: where θ is the initial parameter of the feature encoder based on parallel collaborative attention, θ i ″ is the parameter updated after inner loop training, δ is the training control variable, α and β are the inner loop and outer loop learning rates respectively, L Ti It's a mission loss.
4. The few-sample cross-domain bearing fault diagnosis method based on improved model-independent meta-learning according to claim 3, Features: The output of the two-dimensional convolution module in step 3 is as follows: F=Maxpooling(BN(ReLU(Conv(x)))) (4) Among them, F is the output of the two-dimensional convolution module, x is the input of the two-dimensional convolution module, Maxpooling() is the maximum pooling operation, BN() is batch normalization, ReLU is the activation function, and Conv() is the convolution operation.
5. According to claim 4, the few-sample cross-domain bearing fault diagnosis method based on improved model-independent meta-learning, Features: The output of the parallel collaborative attention module in step 3 is as follows: in, is the feature map that aggregates the other two spatial dimensions, are two channel feature statistics, μ and ω are trainable floating parameters, is the attention weight, F′ W , F′ H , F′ C is an enhanced feature map in three dimensions: width, height, and channel, and F′ is the final feature map output by the parallel collaborative attention module.
6. The few-sample cross-domain bearing fault diagnosis method based on improved model-independent meta-learning according to claim 5, Features: The method updates the network parameters through the Adam optimizer, specifically: Among them, θ t is the network parameter at time t, θ t+1 is the network parameter at time t+1, α is the learning rate, is the first-order moment estimate of the gradient, is the second moment estimate of the gradient, and ε is a small positive value.
7. A few-sample cross-domain bearing fault diagnosis system based on improved model-independent meta-learning, Features: The system includes: a data acquisition and extraction module, a data processing and partitioning module, a feature encoder building module, an improved model-independent meta-learning training module, a testing module, and a fault state identification module; The data acquisition and extraction module is used to extract a small amount of samples, namely the original vibration signal of the bearing; The data processing and partitioning module is used to calculate the self-similarity of the original vibration signal and convert it into a two-dimensional recursive graph, which is divided into a support set and a query set; The feature encoder building module is used to build a feature encoder based on parallel collaborative attention; The improved model-independent meta-learning training module is used to train a feature encoder based on parallel collaborative attention using a support set and a query set to obtain a trained feature encoder; The testing module is used to test the trained parallel collaborative attention-based feature encoder using cross-domain data under other working conditions; The fault status identification module is used to identify the original vibration signal of the bearing and diagnose the bearing fault type.
