Fault migration diagnosis method based on multi-information reconstruction attention mechanism
By adopting a multi-information reconstruction attention mechanism in troubleshooting, the problem of incomplete information learning of the channel attention mechanism is solved, which significantly improves the accuracy and robustness of fault diagnosis.
Patent Information
- Application Number
- CN202411293613.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-14
- Publication Date
- 2025-02-07
AI Technical Summary
In the prior art, when dealing with fault diagnosis in complex industrial systems, the channel attention mechanism is susceptible to interference from redundant information, resulting in a decrease in diagnostic accuracy.
Using the fault migration diagnosis method based on the multi-information reconstruction attention mechanism, the data set is constructed through the sliding window method, and feature information is extracted using multi-scale convolution parallel branches, and input it into the improved channel attention mechanism to reduce redundancy The impact of information enhances the model's ability to identify faults.
It significantly improves the model's ability to identify faults, reduces the impact of redundant information, and improves the accuracy and robustness of fault diagnosis.
Smart Images
Figure CN119397323A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of industrial system fault diagnosis, and specifically relates to a fault migration diagnosis method based on a multi-information reconstruction attention mechanism. Background Art
[0002] With the development of industrialization and the advancement of science and technology, the structure of mechanical equipment is becoming more and more complex, the functions are becoming more and more perfect, and the degree of automation is becoming higher and higher. However, mechanical equipment often faces failure problems during production and operation. If the failure cannot be diagnosed in time, it will lead to a decrease in production efficiency and may even endanger life safety. Therefore, it is particularly important to propose an accurate and feasible fault diagnosis method. Deep learning has been widely used in many fields, and the channel attention mechanism, as a module to enhance the learning ability of the model, is also continuously improved and innovated. Summary of the invention
[0003] The purpose of the present invention is to provide a fault migration diagnosis method based on a multi-information reconstruction attention mechanism, which solves the problem of incomplete information observation of the channel attention mechanism in the prior art. When dealing with fault diagnosis in complex industrial systems, the existing channel attention mechanism is often interfered by redundant information, resulting in a decrease in diagnostic accuracy.
[0004] The technical solution adopted by the present invention is a fault migration diagnosis method based on a multi-information reconstruction attention mechanism, which is specifically implemented in the following steps:
[0005] Step 1: Use the sliding window method to construct the data set;
[0006] Step 2: Use a feature extractor to extract feature information from the input data;
[0007] Step 3: Input the feature information extracted in step 2 into the improved channel attention mechanism to extract the model attention information;
[0008] Step 4: Input the results of the attention mechanism into the classifier and domain classifier respectively to realize the diagnosis of the fault.
[0009] The present invention is also characterized in that:
[0010] Step 1 is implemented as follows:
[0011] Under variable working conditions, vibration acceleration data is collected and classified into normal state and fault state. Due to the change of load, the data distribution of fault signals will be different. Therefore, four fault signals under different loads are selected, which are 0HP, 1HP, 2HP, and 3HP. HP represents different loads input to the motor, and then a data set containing 4 different working conditions is constructed. The data is divided into samples of 2048 data points by the sliding window method. Each sample set contains 100 samples. The samples in each sample set contain normal state and fault state. After the data set is constructed, the entire data set is divided into training set and validation set in a ratio of 7:3.
[0012] Step 2 is implemented as follows:
[0013] The training set constructed in step 1 is proportionally input into the feature extractor for feature extraction. The feature extractor is a multi-scale convolution parallel branch, including three parallel branches, each of which consists of a convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer. Specifically, the convolution kernel sizes of the three branches are 1×3, 1×5, and 1×7, respectively, and the padding is set to 1. The step size of the pooling layer is set to 2, and the window size is 2. After the information extracted by the convolution kernels of the three branches is fused, it is input into four consecutive convolution modules in series. Each module contains a convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer. The convolution kernel size of each convolution module is 1×3, the padding is set to 1, and the step size and window size of the pooling layer are set to 2. The output size of each convolution layer is as follows: the output of the first layer is (32,32,32), the output of the second layer is (32,64,16), the output of the third layer is (32,128,8), and the output of the fourth layer is (32,256,1);
[0014] After the input training set is processed by the multi-scale convolution parallel branch, it is input into the serial convolution module for deeper information extraction to obtain the fault feature information in the bearing signal. The combination of the multi-scale convolution parallel branch and the serial convolution module enables the feature extractor to extract the features of the input signal from different scales and depths, thereby improving the accuracy and robustness of fault diagnosis.
[0015] Step 3 is implemented as follows:
[0016] The feature information extracted by the feature extractor in step 2 is input into the improved attention mechanism MIRCA to enhance the model's recognition effect on faults. The improved attention mechanism MIRCA The overall implementation process of MIRCA includes the following parts. First, the input feature information is passed through the global average pooling layer GAP and the global maximum pooling layer GMP to compress the feature information. Then, a 1*5 convolution kernel is used to learn the correlation between each channel in the compressed feature information. The convolution result is passed through the sigmoid function to divide the correlation range of each channel between 0 and 1. Then, it is sorted according to the size of the correlation, and the feature information is divided into two parts with large correlation and small correlation according to the size of the channel. The two parts of the feature information are expanded through the fully connected layer. The number of channels of the expanded information becomes the number of channels of the original feature information. The 1*5 convolution kernel is used to learn the channel relationship of the expanded information respectively, and the results are added. Then, the 1*5 convolution kernel is used to extract the channel relationship. Finally, the convolution result is passed through the sigmoid function to divide the range between 0 and 1 to obtain the final weight matrix. The weight matrix is multiplied by the input feature information to obtain the feature information that the model pays attention to.
[0017] Step 4 is implemented according to the following steps:
[0018] In step 3, the output of the attention mechanism is flattened and then input into the fault classifier and domain classifier respectively. The fault classifier is used for fault diagnosis. It consists of three fully connected layers, where the number of layers of each fully connected layer is 256, 128, and 10 respectively. Because the target is a 10-classification task, the output of the last layer is set to 10. The domain classifier is used to learn the same features of the source domain and target domain data. The domain classifier consists of three fully connected layers, where the number of layers of each fully connected layer is 128, 64, and 2 respectively, where 2 represents the type of the source domain and target domain of the training sample. The domain classifier is used to reduce the difference in data distribution between different domains and increase the fault diagnosis accuracy of the model.
[0019] The beneficial effect of the present invention is that the fault migration diagnosis method based on the multi-information reconstruction attention mechanism significantly enhances the multi-angle, multi-faceted and multi-level learning ability of the channel attention mechanism for feature information by introducing different types of feature information and adopting convolution kernels of different scales. Specifically, this method can focus on the extracted feature information, increase the weight of the useful information for model training, and reduce the influence of redundant information, thereby significantly improving the model's ability to identify faults. In order to verify the effectiveness of the proposed channel attention mechanism, experimental verification was carried out on the CWRU public dataset. The experimental results show that this method has achieved high-precision detection results in fault detection, proving its effectiveness and reliability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 It is a model structure diagram of the fault migration diagnosis method based on the multi-information reconstruction attention mechanism of the present invention;
[0021] Figure 2 is a structural diagram of a feature extractor in the present invention;
[0022] Figure 3 It is a structural diagram of the multi-information reconstruction attention mechanism in the present invention;
[0023] Figure 4 is a curve chart showing the change in loss of the model in the training set in the present invention;
[0024] Figure 5 It is a curve chart showing the change in accuracy of the model in the validation set in the present invention. DETAILED DESCRIPTION
[0025] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0026] Aiming at the problem that the channel attention mechanism does not fully learn information due to the excessive impact of redundant information on the model learning ability in deep learning, this paper proposes a multi-information reconstruction channel attention mechanism called MIRCA (Multi-information Reconstruction Channel Attention) to increase the network's ability to recognize fault information. The structural model is as follows Figure 3 As shown. Specifically, the feature information is first compressed through different pooling layers, and the compressed information is combined, and the channel relationship of the feature information is preliminarily extracted by convolution. Since the convolution result can reflect the size of the relationship between each channel, the extracted results are sorted from high to low. The sorted information is divided into two parts, and the channel information is expanded through the fully connected layer to become the number of channels of the original feature information. Then, the two expanded feature information are convoluted respectively, the results are added, and the channel relationship between the feature information is learned again by convolution, so as to obtain the important feature information that the model needs to pay attention to. The MIRCA attention mechanism module can improve the model's attention to key information in the fault, reduce the impact of redundant information, and thus increase the model's fault diagnosis ability.
[0027] The present invention is based on a fault migration diagnosis method based on a multi-information reconstruction attention mechanism, combined with Figure 1 to Figure 3 , follow the steps below:
[0028] Step 1: Use the sliding window method to construct the data set;
[0029] Step 1 is implemented as follows:
[0030] Under variable working conditions, vibration acceleration data is collected and classified into normal state and fault state. Fault signals under four different loads are selected, which are 0HP, 1HP, 2HP, and 3HP. HP represents different loads input to the motor. Then, a data set containing four different working conditions is constructed. The data is divided into samples of 2048 data points by the sliding window method. Each sample set contains 100 samples. The samples in each sample set contain normal state and fault state to ensure data diversity and representativeness. After the data set is constructed, the entire data set is divided into a training set and a validation set in a ratio of 7:3. The training set is used for the training process of the model, while the validation set is used to evaluate the performance of the model. This division method helps to ensure the generalization ability of the model under different working conditions, thereby improving the accuracy and reliability of fault diagnosis.
[0031] Step 2: Use a feature extractor to extract feature information from the input data;
[0032] Step 2 is implemented as follows:
[0033] The training set constructed in step 1 is input into the feature extractor in proportion for feature extraction. The structure of the feature extractor is as follows: Figure 2 As shown in the figure. The feature extractor is a multi-scale convolution parallel branch, which includes three parallel branches. Each branch consists of a convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer. Specifically, the convolution kernel sizes of the three branches are 1×3, 1×5, and 1×7, respectively, and the padding is set to 1 to ensure that the boundary information of the input signal is not lost. The step size of the pooling layer is set to 2, and the window size is 2, which is used to reduce the size of the feature map while retaining important feature information. By using convolution kernels of different scales, preliminary information of the input signal can be extracted from different angles, thereby improving the diversity and robustness of feature extraction. After the information extracted by the convolution kernels of the three branches is fused, it is input into four consecutive convolution modules in series. Each module contains a convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer. The convolution kernel size of each convolution module is 1×3, and the padding is set to 1 to ensure that the size of the feature map remains unchanged. The step size and window size of the pooling layer are both set to 2 to further reduce the size of the feature map while retaining important feature information. The output size of each convolutional layer is as follows: the output of the first layer is (32,32,32), the output of the second layer is (32,64,16), the output of the third layer is (32,128,8), and the output of the fourth layer is (32,256,1);
[0034] After the input training set is processed by the multi-scale convolution parallel branch, it will be input into the serial convolution module for deeper information extraction, thereby obtaining a richer and more abstract feature representation. In summary, step 2 uses the feature extractor to extract features from the training set to obtain fault feature information in the bearing signal. The combination of the multi-scale convolution parallel branch and the serial convolution module enables the feature extractor to extract features of the input signal from different scales and depths, thereby improving the accuracy and robustness of fault diagnosis.
[0035] Step 3: Input the feature information extracted in step 2 into the improved channel attention mechanism to extract the model attention information;
[0036] Step 3 is implemented as follows:
[0037] The feature information extracted by the feature extractor in step 2 is input into the improved attention mechanism MIRCA to enhance the model's fault recognition effect. The structure of the improved attention mechanism MIRCA is as follows: Figure 3 As shown in the figure, the overall implementation process of MIRCA includes the following parts. First, the input feature information is passed through the global average pooling layer GAP (Global AveragePooling) and the global maximum pooling layer GMP (Global Max Pooling) to compress the feature information. Then, a 1*5 convolution kernel is used to learn the correlation between each channel in the compressed feature information. The convolution result is passed through the sigmoid function to divide the correlation range of each channel between 0 and 1. Then, it is sorted according to the size of the correlation, and the feature information is divided into two parts with large correlation and small correlation according to the size of the channel. The two parts of feature information are expanded through the fully connected layer. The number of channels of the expanded information becomes the number of channels of the original feature information. The 1*5 convolution kernel is used to learn the channel relationship of the expanded information respectively, and the results are added. Then, the 1*5 convolution kernel is used to extract the channel relationship. Finally, the convolution result is passed through the sigmoid function to divide the range between 0 and 1 to obtain the final weight matrix. The weight matrix is multiplied by the input feature information to obtain the feature information that the model focuses on.
[0038] Step 4: Input the results of the attention mechanism into the classifier and domain classifier respectively to realize the diagnosis of the fault.
[0039] Step 4 is implemented according to the following steps:
[0040] In step 3, the output result of the attention mechanism is flattened and then input into the fault classifier and domain classifier respectively. The fault classifier is used for fault diagnosis. The fault classifier consists of three fully connected layers, where the number of layers of each fully connected layer is 256, 128, and 10 respectively. Because the target is a 10-classification task, the output of the last layer is set to 10. The domain classifier is used to learn the same features of the source domain and target domain data. The domain classifier consists of three fully connected layers, where the number of layers of each fully connected layer is 128, 64, and 2 respectively, where 2 represents the type of the source domain and target domain of the training sample. The domain classifier is used to reduce the difference in data distribution between different domains, improve the recognition ability of the model under variable working condition tasks, and increase the fault diagnosis accuracy of the model. By introducing a multi-information reconstruction attention mechanism, the present invention can observe and analyze feature information more comprehensively, thereby effectively reducing the impact of redundant information.
[0041] Construct evaluation indicators to judge the diagnosis results of the fault diagnosis model established in step 4. The specific implementation is as follows:
[0042] In step 1, the training set and validation set are divided. In order to illustrate the effect of the model, the accuracy rate Acc (Accuracy) is used as the evaluation indicator of the model result. The formula is as follows:
[0043]
[0044] Where Acc represents the ratio of correctly predicted samples to the total number of samples, TP represents the number of samples whose prediction results are normal and whose labels are also normal; FP represents the number of samples whose prediction results are normal but whose labels are abnormal; FN represents the number of samples whose prediction results are abnormal but whose labels are normal; TN represents the number of samples whose prediction results are abnormal and whose labels are also abnormal.
[0045] Example 1
[0046] The fault migration diagnosis method based on the multi-information reconstruction attention mechanism of the present invention is specifically implemented according to the following steps:
[0047] Step 1: Use the sliding window method to construct the data set;
[0048] Step 2: Use a feature extractor to extract feature information from the input data;
[0049] Step 3: Input the feature information extracted in step 2 into the improved channel attention mechanism to extract the model attention information;
[0050] Step 4: Input the results of the attention mechanism into the classifier and domain classifier respectively to realize the diagnosis of the fault.
[0051] Example 2
[0052] The fault migration diagnosis method based on the multi-information reconstruction attention mechanism of the present invention is specifically implemented according to the following steps:
[0053] Step 1: Use the sliding window method to construct the data set;
[0054] Step 1 is implemented as follows:
[0055] Under variable working conditions, vibration acceleration data is collected and classified into normal state and fault state. Fault signals under four different loads are selected, which are 0HP, 1HP, 2HP, and 3HP. HP represents different loads input to the motor. Then, a data set containing four different working conditions is constructed. The data is divided into samples of 2048 data points by the sliding window method. Each sample set contains 100 samples. The samples in each sample set contain normal state and fault state to ensure data diversity and representativeness. After the data set is constructed, the entire data set is divided into a training set and a validation set in a ratio of 7:3. The training set is used for the training process of the model, while the validation set is used to evaluate the performance of the model. This division method helps to ensure the generalization ability of the model under different working conditions, thereby improving the accuracy and reliability of fault diagnosis.
[0056] Step 2: Use a feature extractor to extract feature information from the input data;
[0057] Step 2 is implemented as follows:
[0058] The training set constructed in step 1 is input into the feature extractor in proportion for feature extraction. The structure of the feature extractor is as follows: Figure 2As shown in the figure. The feature extractor is a multi-scale convolution parallel branch, which includes three parallel branches. Each branch consists of a convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer. Specifically, the convolution kernel sizes of the three branches are 1×3, 1×5, and 1×7, respectively, and the padding is set to 1 to ensure that the boundary information of the input signal is not lost. The step size of the pooling layer is set to 2, and the window size is 2, which is used to reduce the size of the feature map while retaining important feature information. By using convolution kernels of different scales, preliminary information of the input signal can be extracted from different angles, thereby improving the diversity and robustness of feature extraction. After the information extracted by the convolution kernels of the three branches is fused, it is input into four consecutive convolution modules in series. Each module contains a convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer. The convolution kernel size of each convolution module is 1×3, and the padding is set to 1 to ensure that the size of the feature map remains unchanged. The step size and window size of the pooling layer are both set to 2 to further reduce the size of the feature map while retaining important feature information. The output size of each convolutional layer is as follows: the output of the first layer is (32,32,32), the output of the second layer is (32,64,16), the output of the third layer is (32,128,8), and the output of the fourth layer is (32,256,1);
[0059] After the input training set is processed by the multi-scale convolution parallel branch, it will be input into the serial convolution module for deeper information extraction, thereby obtaining a richer and more abstract feature representation. In summary, step 2 uses the feature extractor to extract features from the training set to obtain fault feature information in the bearing signal. The combination of the multi-scale convolution parallel branch and the serial convolution module enables the feature extractor to extract features of the input signal from different scales and depths, thereby improving the accuracy and robustness of fault diagnosis.
[0060] Step 3: Input the feature information extracted in step 2 into the improved channel attention mechanism to extract the model attention information;
[0061] Step 4: Input the results of the attention mechanism into the classifier and domain classifier respectively to realize the diagnosis of the fault.
[0062] Step 4 is implemented according to the following steps:
[0063] In step 3, the output of the attention mechanism is flattened and then input into the fault classifier and domain classifier respectively. The fault classifier is used for fault diagnosis. The fault classifier consists of three fully connected layers, where the number of layers of each fully connected layer is 256, 128, and 10 respectively. Because the target is a 10-classification task, the output of the last layer is set to 10. The domain classifier is used to learn the same features of the source domain and target domain data. The domain classifier consists of three fully connected layers, where the number of layers of each fully connected layer is 128, 64, and 2 respectively, where 2 represents the type of the source domain and target domain of the training sample. The domain classifier is used to reduce the difference in data distribution between different domains, improve the recognition ability of the model under variable working condition tasks, and increase the fault diagnosis accuracy of the model.
[0064] Construct evaluation indicators to judge the diagnosis results of the fault diagnosis model established in step 4. The specific implementation is as follows:
[0065] In step 1, the training set and validation set are divided. In order to illustrate the effect of the model, the accuracy rate Acc (Accuracy) is used as the evaluation indicator of the model result. The formula is as follows:
[0066]
[0067] Where Acc represents the ratio of correctly predicted samples to the total number of samples, TP represents the number of samples whose prediction results are normal and whose labels are also normal; FP represents the number of samples whose prediction results are normal but whose labels are abnormal; FN represents the number of samples whose prediction results are abnormal but whose labels are normal; TN represents the number of samples whose prediction results are abnormal and whose labels are also abnormal.
[0068] Example 3
[0069] The present invention uses the rolling bearing fault data set published by Western Reserve University to verify the effectiveness and reliability of the proposed method under different working conditions. The data set contains fault data of the drive end and fan end bearings, of which the drive end has two sets of data with sampling frequencies of 12kHz and 48kHz, respectively, while the fan end has only one set of data with a sampling frequency of 12kHz. In order to conduct a detailed analysis, the present invention selects the drive end bearing data with a sampling frequency of 12kHz. The data set covers four different health states of rolling bearings: normal state, inner ring fault, outer ring fault and rolling element fault. In addition, the data set also includes three faults of different sizes, namely 7mils, 14mils and 21mils, and is tested under four different working conditions, namely: 0HP, 1HP, 2HP and 3HP. In order to verify the effectiveness of the proposed channel attention mechanism, the present invention selects working conditions with different source domains and target domains for experimental verification. Specifically, multiple channel attention mechanisms that are currently widely used are selected for effect comparison, including channel attention module (CAM) and efficient channel attention (ECA). The experimental results are shown in Table 1, showing the performance of different attention mechanisms in fault identification. Through these experiments, the present invention proves the effectiveness and reliability of the proposed method under different working conditions, and demonstrates its potential in rolling bearing fault diagnosis. The experimental results show that the proposed channel attention mechanism can significantly improve the model's ability to identify faults, reduce the impact of redundant information, and thus improve detection accuracy. Table 1 Effect comparison of different attention mechanisms Source domain Target domain Original CAMECA This method 0HP 1HP 95.67% 97.00% 99.33% 99.67% 1HP 2HP 99.67% 99.67% 99.67% 99.67% 99.67% 2HP 3HP 98.33% 99.33% 98.67% 99.67% Average 97.67% 97.61% 98.33% 99.17% It can be seen from Table 1 that the proposed attention mechanism can provide effective help for model learning in fault diagnosis under variable working conditions. From the feature information learning of the source domain with 1HP to the target domain with 3HP, it can be clearly seen that the accuracy of CAM and ECA has decreased, but the proposed method can be improved to 2.34% compared with the original model. Figure 4 It can be seen that in the training set, the training loss curve can converge around 160 generations, and finally the training set loss tends to 0. Figure 5 The accuracy change curve of the validation set shows that the model can reach stability in about 60 generations, and the accuracy approaches 100%.
Claims
1. A fault migration diagnosis method based on multi-information reconstruction attention mechanism, characterized in that: Follow the steps below to implement it: Step 1: Use the sliding window method to construct the data set; Step 2: Use a feature extractor to extract feature information from the input data; Step 3: Input the feature information extracted in step 2 into the improved channel attention mechanism to extract the model attention information; Step 4: Input the results of the attention mechanism into the classifier and domain classifier respectively to realize the diagnosis of the fault.
2. The fault migration diagnosis method based on multi-information reconstruction attention mechanism according to claim 1 is characterized in that: The step 1 is specifically implemented according to the following steps: Under variable working conditions, vibration acceleration data is collected and classified into normal state and fault state. Fault signals under four different loads are selected, and the fault signals under four different loads are 0HP, 1HP, 2HP, and 3HP, respectively. HP represents different loads input to the motor. Then, a data set containing four different working conditions is constructed. The data is divided into samples of 2048 data points through the sliding window method. Each sample set contains 100 samples, and the samples in each sample set contain normal state and fault state. After the data set is constructed, the entire data set is divided into training set and validation set in a ratio of 7:
3.
3. The fault migration diagnosis method based on multi-information reconstruction attention mechanism according to claim 2 is characterized in that: The step 2 is specifically implemented according to the following steps: The training set constructed in step 1 is proportionally input into the feature extractor for feature extraction. The feature extractor is a multi-scale convolution parallel branch, including three parallel branches, each of which consists of a convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer. Specifically, the convolution kernel sizes of the three branches are 1×3, 1×5, and 1×7, respectively, and the padding is set to 1. The step size of the pooling layer is set to 2, and the window size is 2. After the information extracted by the convolution kernels of the three branches is fused, it is input into four consecutive convolution modules in series. Each module contains a convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer. The convolution kernel size of each convolution module is 1×3, the padding is set to 1, and the step size and window size of the pooling layer are set to 2. The output size of each convolution layer is as follows: the output of the first layer is (32,32,32), the output of the second layer is (32,64,16), the output of the third layer is (32,128,8), and the output of the fourth layer is (32,256,1); After the input training set is processed by the multi-scale convolution parallel branch, it will be input into the serial convolution module for deeper information extraction to obtain the fault feature information in the bearing signal. The combination of the multi-scale convolution parallel branch and the serial convolution module enables the feature extractor to extract features of the input signal from different scales and depths, thereby improving the accuracy and robustness of fault diagnosis.
4. The fault migration diagnosis method based on multi-information reconstruction attention mechanism according to claim 3 is characterized in that: The step 3 is specifically implemented according to the following steps: The feature information extracted by the feature extractor in step 2 is input into the improved attention mechanism MIRCA to enhance the model's recognition effect on faults. The improved attention mechanism MIRCA The overall implementation process of MIRCA includes the following parts. First, the input feature information is passed through the global average pooling layer GAP and the global maximum pooling layer GMP to compress the feature information. Then, a 1*5 convolution kernel is used to learn the correlation between each channel in the compressed feature information. The convolution result is passed through the sigmoid function to divide the correlation range of each channel between 0 and 1. Then, it is sorted according to the size of the correlation, and the feature information is divided into two parts with large correlation and small correlation according to the size of the channel. The two parts of the feature information are expanded through the fully connected layer. The number of channels of the expanded information becomes the number of channels of the original feature information. The 1*5 convolution kernel is used to learn the channel relationship of the expanded information respectively, and the results are added. Then, the 1*5 convolution kernel is used to extract the channel relationship. Finally, the convolution result is passed through the sigmoid function to divide the range between 0 and 1 to obtain the final weight matrix. The weight matrix is multiplied by the input feature information to obtain the feature information that the model pays attention to.
5. The fault migration diagnosis method based on multi-information reconstruction attention mechanism according to claim 4 is characterized in that: The step 4 is specifically implemented according to the following steps: In step 3, the output results of the attention mechanism are flattened and then input into the fault classifier and domain classifier respectively. The fault classifier is used for fault diagnosis. The fault classifier consists of three fully connected layers, where the number of layers of each fully connected layer is 256, 128, and 10 respectively. Because the goal is a 10-classification task, the output of the last layer is set to 10. The domain classifier is used to learn the same features of the source domain and target domain data. The domain classifier consists of three fully connected layers, where the number of layers of each fully connected layer is 128, 64, and 2 respectively, where 2 represents the type of the source domain and target domain of the training sample. The domain classifier is used to reduce the difference in data distribution between different domains and increase the fault diagnosis accuracy of the model.