Bearing fault diagnosis method based on parallel multi-fusion deep learning network model

Abstract

The invention discloses a bearing fault diagnosis method based on a parallel multi-fusion deep learning network model. And obtaining a sample data set, dividing the sample data set into a training data set and a test data set, carrying out parallel multi-fusion deep learning network model training by using the training data set, and inputting the test data set into the trained parallel multi-fusion deep learning network model to carry out fault diagnosis. Wherein the parallel multi-fusion deep learning network model comprises three parallel sub-networks, namely a bidirectional gating circulation unit sub-network, a fast Fourier transform-convolutional neural network sub-network and a discrete wavelet transform-enhanced depth separable ion network, and a dynamic multi-head attention mechanism module. According to the method, the three parallel sub-networks and the dynamic multi-head attention mechanism module are combined to automatically extract and identify the fault features of the rolling bearing under multiple working conditions, and self-adaptive fault feature extraction under different domain signals and rolling bearing fault diagnosis under complex working conditions are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent diagnosis of bearing faults, and specifically to a bearing fault diagnosis method based on a parallel multi-fusion deep learning network model. Background Art

[0002] As a key component in mechanical equipment, rolling bearings often operate at high speed, heavy load, high temperature and changeable working environment, so they are prone to failure. According to statistics, about 30% to 40% of mechanical failures are caused by bearings, which not only bring safety hazards but also economic losses. Because vibration signals can reflect rich status information, fault diagnosis technology based on vibration signals has been widely used in the field of bearing fault diagnosis. With the development of artificial intelligence technology, intelligent fault diagnosis methods have attracted the attention of many researchers. The traditional intelligent fault diagnosis process usually includes the following steps: first, collect and preprocess the vibration signal; then extract features from the processed signal; finally, use a machine learning-based classifier to identify the health status of the bearing. However, the feature extraction in this traditional method relies on the experience of experts and advanced signal processing technology, which makes it very difficult to select a set of suitable features, and machine learning cannot mine deep features. Deep learning, an end-to-end fault diagnosis method, has been introduced into bearing fault diagnosis because it can automatically learn effective feature representations from raw data. However, most of the existing methods focus on combining the input of a single feature extraction with a single network, and most of the above methods ignore the temporal characteristics of the original vibration signal. Only the features of a single domain can be extracted separately, which makes the feature extraction incomplete and cannot fully reflect the fault feature information. In actual engineering, due to the influence of operating conditions and environment, the collected vibration signals often show different characteristics in different domains. Relying solely on the features of a single domain may miss some information, thus affecting the final diagnosis accuracy. Summary of the invention

[0003] In order to address the deficiencies in the prior art, the present invention provides a bearing fault diagnosis method based on a parallel multi-fusion deep learning network model. By combining a bidirectional gated recurrent unit subnetwork, a fast Fourier transform-convolutional neural network subnetwork, a discrete wavelet transform-enhanced deep separable subnetwork and a dynamic multi-head attention mechanism module, the fault features of rolling bearings under multiple working conditions are automatically extracted and identified, thereby realizing adaptive fault feature extraction under different domain signals and rolling bearing fault diagnosis under complex working conditions.

[0004] In order to achieve the above object, the specific scheme adopted by the present invention is: The bearing fault diagnosis method based on the parallel multi-fusion deep learning network model mainly includes the following steps: Step S1, obtaining vibration signals of the bearing under multiple working conditions through a public data set, labeling and segmenting the collected vibration signals to obtain a sample data set of the vibration signals, wherein the sample data set includes a training data set and a test data set; Step S2, training a parallel multi-fusion deep learning network model using the training data set; Step S3, inputting the test data set into the trained parallel multi-fusion deep learning network model, thereby obtaining the fault diagnosis result of the test data; Among them, in step S2, the parallel multi-fusion deep learning network model includes three parallel sub-networks of a bidirectional gated recurrent unit sub-network, a fast Fourier transform-convolutional neural network sub-network, and a discrete wavelet transform-enhanced deep separable sub-network, and a dynamic multi-head attention mechanism module; the specific method of training the parallel multi-fusion deep learning network model using the training data set is: Step S21, input the training data set into the bidirectional gated recurrent unit sub-network, the fast Fourier transform-convolutional neural network sub-network and the discrete wavelet transform-enhanced deep separable sub-network at the same time, and extract features in the time domain, frequency domain and time-frequency domain respectively; Step S22, input the features obtained in step S21 into the dynamic multi-head attention mechanism module for interactive learning, capture the intrinsic connection and complementary information between different features, and fuse the features of interactive learning; Step S23: Flatten the fused features into a one-dimensional vector and then input them into the Softmax layer, thus completing the feature learning and classification process of the entire model and obtaining a trained parallel multi-fusion deep learning network model.

[0005] Furthermore, in step S2, during the entire training process of the model, the cross entropy loss function is used to calculate the loss of the predicted and true labels, and the back propagation algorithm is used to update the weight parameters.

[0006] Furthermore, step S1 specifically includes the following steps: Step S11, obtain sample data set I: obtain the vibration signals of rolling bearings at the same sampling frequency, different loads, and different speeds through public data sets, the sampling frequency is 25.6KHz, and the rolling bearings include four states: inner ring fault, outer ring fault, rolling element fault, and normal state; take 500 data points as a sample data, and assign sample data labels according to the rolling bearing state, and combine multiple sample data to obtain sample data set I, 70% of the sample data set I is divided out as a training data set, and the remaining 30% is used as a test data set.

[0007] Furthermore, the step S1 specifically further includes the following steps: Step S12, obtain sample data set II: obtain the vibration signal of the rolling bearing at the same sampling frequency and different speeds through the public data set, the sampling frequency is 20KHz, and the rolling bearing includes four states: inner ring fault, outer ring fault, rolling element fault and normal state; take 500 data points as a sample data, and assign sample data labels according to the rolling bearing state, and combine multiple sample data to obtain sample data set II, 70% of the sample data set II is divided out as a training data set, and the remaining 30% is used as a test data set.

[0008] Furthermore, the bidirectional gated recurrent unit sub-network adaptively learns dynamic information based on the directly input original signal and performs time domain feature extraction; The Fast Fourier Transform-Convolutional Neural Network sub-network uses Fast Fourier Transform to transform the original signal into a frequency domain signal, which is then input into the Convolutional Neural Network to extract frequency domain features. The discrete wavelet transform-enhanced deep separable sub-network uses discrete wavelet transform to transform the one-dimensional time series into two-dimensional time-frequency domain image features, and inputs the time-frequency domain image into the enhanced deep separable sub-network to extract time-frequency domain features.

[0009] Furthermore, the fast Fourier transform-convolutional neural network subnetwork includes a fast Fourier transform layer, three convolution layers, and three maximum pooling layers; the signals output by each convolution layer enter the maximum pooling layer for processing, and after the signal is output by the third convolution layer, it enters the dynamic multi-head attention mechanism module after being processed by the maximum pooling layer; The discrete wavelet transform-enhanced deep separable subnetwork includes a discrete wavelet transform layer, three deep separable convolutional layers, and three maximum pooling layers. The signals output by each deep separable convolutional layer enter the maximum pooling layer for processing. After the signal is output by the third deep separable convolutional layer, it enters the dynamic multi-head attention mechanism module through the maximum pooling layer. An SRM high-pass filter is set between the discrete wavelet transform layer and the first depth-wise separable convolutional layer, which makes it easier for the network to learn the key information in the image and enhances the high-frequency information.

[0010] Furthermore, the calculation process of the dynamic multi-head attention mechanism module in step S22 is as follows: First, the input Query (Q), Key (K) and Value (V) are mapped to a new vector space through linear transformation. The dimensions of these vectors are d model , the formula is as follows: Q'=W Q Q,K'=W K K,V'=W V ·V In the formula, Q represents the query vector of the input data, K represents the key vector of the input data, V represents the value vector of the input data, and W Q , W K , W V is the learned weight matrix used to convert the input Q, K, V into a new feature representation; Q′ represents the query vector of the new feature, K′ represents the key vector of the new feature, and V′ represents the value vector of the new feature; Introduce the dynamic adjustment parameter α to dynamically adjust Q and K. The formula is as follows: Where Q″ represents the query vector adjusted by the dynamic adjustment parameters, and K″ represents the key vector adjusted by the dynamic adjustment parameters; then the transformed Q″, K″, V′ are divided into h heads, each head has its own Q, K, V, and the dimension of each head is d k , the formula is as follows: Q i =split_heads(Q”),K i =split_heads(K”),V i =split_heads(V') In the formula, Q i represents the query vector of the i-th head, K i represents the key vector of the i-th head, V i Represents the value vector of the i-th head, i represents the i-th head, 1≤i≤h, split_heads is the split function, and the dimension of each head is i.e. k ; Calculate the attention weight of each head and normalize it. The calculation formula is as follows: In the formula, α i represents the dynamic adjustment parameter of the i-th head; Calculate the input of each self-attention mechanism as follows: Attention i =Attention_Weight i ·V i The output of each head is concatenated together and linearly transformed to get the final output: Output=Concat(Attention1,Attention2,...,Attention h )·W O Where W O is the output linear transformation weight matrix.

[0011] Furthermore, step S3 specifically includes the following steps: Step S31, input the test data set in step S1 into the trained bidirectional gated recurrent unit sub-network, fast Fourier transform-convolutional neural network sub-network and discrete wavelet transform-enhanced deep separable sub-network at the same time to obtain a feature parameter matrix; Step S32, input the feature parameter matrix obtained in step S31 into the trained dynamic multi-head attention mechanism module to perform interactive learning, capture the intrinsic connection and complementary information between different features, and fuse the features of interactive learning, flatten the fused features into a one-dimensional vector, and then input them into the Softmax layer, and output the bearing state category of the identified test data set; Step S33: Compare the identified bearing status category of the test data set with the sample data label of the test data set, and calculate the diagnosis and recognition accuracy of the rolling bearing fault.

[0012] Beneficial effects:

[0013] (1) The bearing fault diagnosis method disclosed in the present invention combines a bidirectional gated recurrent unit subnetwork, a fast Fourier transform-convolutional neural network subnetwork, a discrete wavelet transform-enhanced deep separable subnetwork and a dynamic multi-head attention mechanism module to achieve automatic extraction and identification of rolling bearing fault features under different domain signals, thereby greatly improving the accuracy of rolling bearing fault diagnosis under complex working conditions.

[0014] (2) The bearing fault diagnosis method disclosed in the present invention can adaptively learn dynamic information because it has a bidirectional gated recurrent unit sub-network.

[0015] (3) The bearing fault diagnosis method disclosed in the present invention uses fast Fourier transform to convert the original signal into a frequency domain signal, and uses discrete wavelet transform to convert the one-dimensional time series into a two-dimensional time-frequency domain image feature, enriching the diversity of feature extraction of the parallel multi-fusion deep learning network model.

[0016] (4) In the bearing fault diagnosis method disclosed in the present invention, the fast Fourier transform-convolutional neural network subnetwork includes a fast Fourier transform layer, three convolution layers, and three maximum pooling layers; the discrete wavelet transform-enhanced deep separable subnetwork includes a discrete wavelet transform layer, three deep separable convolution layers, and three maximum pooling layers; wherein, an SRM high-pass filter is provided between the discrete wavelet transform layer and the first deep separable convolution layer, so that the network can more easily learn the key information in the image and enhance the high-frequency information.

[0017] (5) The bearing fault diagnosis method disclosed in the present invention includes a dynamic multi-head attention mechanism module that can interact and fuse the features of the three extracted branches, capture the intrinsic connections and complementary information between different features, and enhance the feature expression capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a bearing fault diagnosis flow chart of the present invention.

[0019] Figure 2 It is a structural diagram of the depth-separable convolutional layer of the present invention.

[0020] Figure 3 It is a structural diagram of the dynamic multi-head attention mechanism module of the present invention.

[0021] Figure 4 It is a structural diagram of the bearing fault diagnosis model of the present invention.

[0022] Figure 5 It is the confusion matrix diagram of sample data set I.

[0023] Figure 6 It is the T-SNE visualization of sample data set I.

[0024] Figure 7 It is the confusion matrix diagram of sample data set II.

[0025] Figure 8 This is the T-SNE visualization of sample data set II.

[0026] Fig. 9 It is a comparison chart of the diagnostic accuracy of sample data set I and the comparison model.

[0027] Fig.10 It is a comparison chart of the diagnostic accuracy of sample data set II and the comparison model. DETAILED DESCRIPTION

[0028] The technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0029] Example 1

[0030] This embodiment discloses a rolling bearing fault diagnosis method based on a parallel multi-fusion deep learning network model. Figure 1-4 Each step is explained in detail.

[0031] Step S1, obtaining a sample data set: obtaining vibration signals of rolling bearings under multiple working conditions through public data sets, labeling and segmenting the collected vibration signals, and obtaining a sample data set of vibration signals, wherein the sample data set includes a training data set and a test data set.

[0032] Step S1 includes the following steps: Step S11, obtain sample data set I: obtain the vibration signals of rolling bearings at the same sampling frequency, different loads, and different speeds through public data sets (rolling bearing data set of Huazhong University of Science and Technology and rolling bearing data set of University of Cincinnati), the sampling frequency is 25.6KHz, and the rolling bearings include four states: inner ring fault, outer ring fault, rolling element fault, and normal state; in this embodiment, sample data set I includes the operating results of the bearing under three different speed conditions, namely 65Hz, 70Hz, and 75Hz; 500 data points are taken as one sample data, and the sample data labels are assigned according to the rolling bearing state, and multiple sample data are combined to obtain sample data set I. The specific situation of sample data set I in this embodiment is shown in Table 1.

[0033] Table 1 Details of sample data set I

[0034] Step S12, obtaining sample data set II: obtaining vibration signals of rolling bearings at the same sampling frequency and different speeds through public data sets, the sampling frequency is 20KHz, and the rolling bearings include four states: inner ring fault, outer ring fault, rolling element fault and normal state; taking 500 data points as one sample data, and assigning sample data labels according to the rolling bearing state, and combining multiple sample data to obtain sample data set II. The specific situation of sample data set II is shown in Table 2:

[0035] Table 2 Details of sample data set II

[0036] Step S13: 70% of the sample data set I and the sample data set II are divided into training data sets, and the remaining 30% are used as test data sets.

[0037] Step S2: Establish a parallel multi-fusion deep learning network model, and use the training data set to train the parallel multi-fusion deep learning network model.

[0038] Step S2 is described in detail below.

[0039] Step S21, establish a parallel multi-fusion deep learning network model, the parallel multi-fusion deep learning network model includes three parallel sub-networks of bidirectional gated recurrent unit sub-network, fast Fourier transform-convolutional neural network sub-network, discrete wavelet transform-enhanced deep separable sub-network, a dynamic multi-head attention mechanism module capable of summarizing the output signals of the three parallel sub-networks, a fully connected layer and a softmax layer. Among them, the fast Fourier transform-convolutional neural network sub-network includes a fast Fourier transform layer, a convolutional layer 1, a maximum pooling layer, a convolutional layer 2, a maximum pooling layer, a convolutional layer 3, and a maximum pooling layer in sequence; the discrete wavelet transform-enhanced deep separable sub-network includes a discrete wavelet transform layer, an SRM filter, a deep separable convolutional layer 1, a maximum pooling layer, a deep separable convolutional layer 2, a maximum pooling layer, a deep separable convolutional layer 3, and a maximum pooling layer in sequence.

[0040] Step S22, input the training data set into the bidirectional gated recurrent unit subnetwork, the fast Fourier transform-convolutional neural network subnetwork and the discrete wavelet transform-enhanced deep separable subnetwork at the same time, and perform feature extraction in the time domain, frequency domain and time-frequency domain. Among them, the bidirectional gated recurrent unit subnetwork adaptively learns dynamic information based on the original signal directly input, and extracts time domain features; the fast Fourier transform-convolutional neural network subnetwork uses fast Fourier transform to convert the original signal into a frequency domain signal, and then inputs it into the convolutional neural network to extract frequency domain features; the discrete wavelet transform-enhanced deep separable subnetwork uses discrete wavelet transform to convert the one-dimensional time series into two-dimensional time-frequency domain image features, and inputs the time-frequency domain image into the enhanced deep separable subnetwork to extract time-frequency domain features. The enhanced deep separable network adds an SRM high-pass filter before the deep separable convolution, so that the network can more easily learn the key information in the image and enhance the high-frequency information. The SRM high-pass filter is: 1st:[-1 1],2nd: 3rd: edge 3×3 : square 3×3 : edge 5×5 : square 5×5 :

[0041] The above 7 high-pass filters can extract high-frequency feature information of 7 different modes. The above 7 high-pass filters are all expanded to a size of 5×5.

[0042] Step S23: Input the features obtained in step S22 into the dynamic multi-head attention mechanism module for interactive learning, capture the intrinsic connection and complementary information between different features, and fuse the features of interactive learning. The scaling factor of the traditional multi-head attention mechanism is fixed, which makes the model lack flexibility. Therefore, a parameter α is introduced. When calculating the attention weight, α is used in the weight calculation to dynamically adjust the scaling factor of each head. The calculation process of the dynamic multi-head attention mechanism is as follows: First, the input Query (Q), Key (K) and Value (V) are mapped to a new vector space through linear transformation. The dimensions of these vectors are d model , the formula is as follows: Q'=W Q Q,K'=W K K,V'=W V ·V In the formula, Q represents the query vector of the input data, K represents the key vector of the input data, V represents the value vector of the input data, and W Q , W K , W V It is the learned weight matrix used to convert the input Q, K, V into a new feature representation; Q′ represents the query vector of the new feature, K′ represents the key vector of the new feature, and V′ represents the value vector of the new feature.

[0043] Then, the dynamic adjustment parameter α is introduced to dynamically adjust Q and K. The formula is as follows: Wherein, Q″ represents the query vector adjusted by the dynamic adjustment parameters, and K″ represents the key vector adjusted by the dynamic adjustment parameters.

[0044] Divide the transformed Q″, K″, V′ into h heads, each with its own Q, K, V, and the dimension of each head is d k , the formula is as follows: Q i =split_heads(Q”),K i =split_heads(K”),V i =split_heads(V') In the formula, Q i represents the query vector of the i-th head, K i represents the key vector of the i-th head, V i Represents the value vector of the i-th head, i represents the i-th head, 1≤i≤h, split_heads is the split function, and the dimension of each head is i.e. k .

[0045] Calculate the attention weight of each head and normalize it. The calculation formula is as follows: In the formula, α i represents the dynamic adjustment parameter of the i-th head;

[0046] Calculate the input of each self-attention mechanism as follows: Attention i =Attention_Weight i ·V i

[0047] The output of each head is concatenated together and linearly transformed to get the final output: Output=Concat(Attention1,Attention2,...,Attention h )·W O

[0048] Where W O is the output linear transformation weight matrix.

[0049] Step S24: Flatten the fused features into a one-dimensional vector and then input them into the Softmax layer to complete the feature learning and classification process of the entire model.

[0050] It should be noted that during the entire training process, the cross entropy loss function is used to calculate the loss of predictions and true labels during model training, and the back propagation algorithm is used to update the weight parameters.

[0051] The number of iterations is 50, the batch size is 32, and the back-propagation algorithm is used to iteratively update the parameters of the bidirectional gated recurrent unit sub-network, the fast Fourier transform-convolutional neural network sub-network, the discrete wavelet transform-enhanced deep separable sub-network and the dynamic multi-head attention mechanism module until the number of iterations is reached, indicating that the parallel multi-fusion deep learning network model training is completed, and step S3 is entered.

[0052] Step S3: input the test data set into the trained parallel multi-fusion deep learning network model to obtain the fault diagnosis result of the test data.

[0053] Step S31, after the parameters of the bidirectional gated recurrent unit subnetwork, the fast Fourier transform-convolutional neural network subnetwork and the discrete wavelet transform-enhanced deep separable subnetwork stop updating, the trained bidirectional gated recurrent unit subnetwork, the fast Fourier transform-convolutional neural network subnetwork and the discrete wavelet transform-enhanced deep separable subnetwork are obtained, and the test data set described in S1 is input into the trained bidirectional gated recurrent unit subnetwork, the fast Fourier transform-convolutional neural network subnetwork and the discrete wavelet transform-enhanced deep separable subnetwork to obtain a feature parameter matrix.

[0054] Step S32: Input the feature parameter matrix obtained in step S31 into the trained dynamic multi-head attention mechanism module for interactive learning to capture the intrinsic connection and complementary information between different features, fuse the features of interactive learning, flatten the fused features into a one-dimensional vector, and then input it into the Softmax layer to output the bearing state category of the identified test data set.

[0055] Step S33: Compare the identified bearing status category of the test data set with the sample data label of the test data set, and calculate the diagnosis and recognition accuracy of the rolling bearing fault.

[0056] The confusion matrix and T-SNE visualization diagram of the sample data set I and sample data set II are as follows Figures 5 to 8 shown. Figure 5 In the confusion matrix, 0-11 correspond to 12 status categories, the values ​​on the diagonal are the recognition accuracy of each status, and the remaining values ​​are the recognition error rates. The results of the confusion matrix show that except for category 7, which has 3.33% misdiagnosed as category 4, the diagnosis rates of the remaining categories are 100%, with no misdiagnosis. Figure 7 The results of the confusion matrix show that the diagnosis rate of all categories is 100%, with no misdiagnosis. Figure 6 and Figure 8 The feature visualization of T-SNE shows that the boundaries of each category are clear. After reducing the dimension from high-dimensional features to a two-dimensional view, each fault category is still highly distinguishable. It verifies that the method proposed in the present invention has demonstrated extremely high classification accuracy and stability in the field of rolling bearing fault identification, as well as the ability to effectively extract and maintain classification differences in complex feature spaces, effectively reducing confusion between categories.

[0057] In order to verify the performance of the method proposed in the present invention, the method of the present invention is compared with traditional CNN, MSCNN, AlexNet and advanced TSCNN, MSTCNN network models. Among them, (1) CNN, MSCNN, AlexNet three models are directly input with the original vibration signal; (2) TSCNN: a parallel network model with two different domains input, time domain and time-frequency domain; (3) MSTCNN: a parallel network model with three different domains input, time domain, frequency domain, and time-frequency domain. To ensure fairness, all models are trained using the same initial parameters as the method proposed in the present invention. The evaluation criteria are accuracy, F1 score, precision and recall. In order to reduce the impact of randomness, each experiment is repeated three times.

[0058] First, we use data set I for verification and obtain the diagnostic accuracy of each model as shown in Table 3 and Fig. 9 shown.

[0059] Table 3 Diagnostic rate results of each model when using data set I Model Average F1 score / % Average accuracy / % Average recall / % CNN 78.73 81.65 79.26 MSCNN 86.17 87.87 86.48 AlexNet 98.61 98.73 98.61 TSCNN 69.69 80.23 75.09 MSTCNN 39.28 47.56 49.63 Model of the present invention 99.63 99.64 99.63

[0060] From Table 3 and Fig. 9 It can be seen that the model proposed in the present invention shows excellent performance that exceeds the comparison model in four evaluation indicators (average F1 score, average precision, average recall rate, and average accuracy). Among the six models involved in the comparison, only the MSTCNN model failed to break through the 50% threshold in performance, while the other models reached a level of more than 70% in all indicators except the average F1 score of TSCNN. Compared with the other five models, the model proposed in the present invention has achieved significant improvements in the average diagnostic accuracy, which is specifically manifested as follows: relative to the comparison model, in terms of average diagnostic rate, the model proposed in the present invention achieved 20.37%, 13.15%, 1.02%, 24.54% and 50% performance improvements respectively. This result proves that the method proposed in the present invention not only performs well in effectiveness, but also has significant advantages in generalization ability and robustness.

[0061] Then we used Dataset II for further verification. The diagnostic rate results of each model are shown in Table 4 and Fig.10 shown.

[0062] Table 4 Diagnostic rate results of each model when using data set II Model Average F1 score / % Average accuracy / % Average recall / % CNN 94.21 95.75 94.54 MSCNN 89.73 92.72 90.16 AlexNet 95.30 96.00 95.44 TSCNN 94.90 96.20 94.90 MSTCNN 80.40 77.82 84.70 Model of the present invention 100 100 100

[0063] From Table 4 and Fig.10It can be seen that among the six models involved in the comparison, all indicators except the average accuracy of MSTCNN reached a level of more than 80%. Compared with the other five models, the model proposed in the present invention has achieved significant improvements in the average diagnostic accuracy. Specifically, compared with the comparison model, in terms of the average diagnostic rate, the model proposed in the present invention achieved a performance improvement of 5.46%, 9.84%, 4.56%, 5.10% and 15.30%, respectively. This result proves that the method proposed in the present invention not only excels in effectiveness, but also has significant advantages in generalization ability and robustness. At the same time Fig.10 It can be seen intuitively that the stability of the method proposed in the present invention is significantly improved.

[0064] In order to verify the effectiveness of each branch in the proposed method, an ablation experiment was conducted. The models for comparative analysis are as follows: (a) no time domain branch (PMFDL w / o T); (b) no frequency domain branch (PMFDL w / o F); (c) no time-frequency domain branch (PMFDL w / o TF); (d) no dynamic multi-head attention mechanism module (PMFDL w / o D); (e) only the time domain branch (OTB); (f) only the frequency domain branch (OFB); (g) only the time-frequency domain branch (OTFB); (h) the proposed method. The settings of the network parameters are the same as those of the proposed method. The results of Dataset I are shown in Table 5.

[0065] Table 5 Diagnostic rate results of each model when using data set I

[0066] It can be seen from Table 5 that the results of the four evaluation indicators of the method proposed in the present invention all exceed 99%. When any one of the branches is removed, the accuracy of the results is significantly reduced. This is because the proposed method can extract multiple features such as time, frequency, and local features of the vibration signal, thereby improving the robustness and generalization of the model. This result proves the indispensability of each branch in the proposed method. And when all branches act alone, the results are also significantly reduced. In particular, the result of the time-frequency domain branch did not exceed 65%, proving the synergistic interaction between the various branches.

[0067] Then, Dataset II was selected for further verification, and the diagnostic accuracy is shown in Table 6.

[0068] Table 6 Diagnostic rate results of each model when using data set II Model Average accuracy / % Average F1 score / % Average accuracy / % Average recall / % PMFDL w / o T 96.90 96.86 97.42 96.90 PMFDL w / o F 93.99 93.64 94.88 93.99 PMFDL w / o TF 96.17 96.10 96.88 96.17 PMFDL w / o D 96.36 96.35 96.58 96.36 OTB 93.80 93.66 94.40 93.80 OFB 91.44 90.97 94.32 91.44 OTFB 93.61 93.67 94.01 93.62 Model of the present invention 100 100 100 100

[0069] As can be seen from Table 6, the results of the four evaluation indicators of the proposed method reached 100%. When any of the branches is removed, the accuracy of the results is above 90%, but it is significantly lower than that of the proposed method. This result proves the indispensability of each branch in the proposed method. And when all branches act alone, the results are also significantly lower.

[0070] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any equivalent changes or modifications made according to the essence of the present invention should be included in the protection scope of the present invention.

Claims

1. Bearing fault diagnosis method based on parallel multi-fusion deep learning network model, It is characterized in that The main steps are as follows: Step S1, obtaining vibration signals of the bearing under multiple working conditions through public data sets, labeling and segmenting the collected vibration signals to obtain sample data sets of the vibration signals, wherein the sample data sets include a training data set and a test data set; Step S2: training the parallel multi-fusion deep learning network model using the training data set; Step S3, inputting the test data set into the trained parallel multi-fusion deep learning network model, thereby obtaining the fault diagnosis result of the test data; Among them, in step S2, the parallel multi-fusion deep learning network model includes three parallel sub-networks of a bidirectional gated recurrent unit sub-network, a fast Fourier transform-convolutional neural network sub-network, and a discrete wavelet transform-enhanced deep separable sub-network, and a dynamic multi-head attention mechanism module; the specific method of training the parallel multi-fusion deep learning network model using the training data set is: Step S21, input the training data set into the bidirectional gated recurrent unit sub-network, the fast Fourier transform-convolutional neural network sub-network and the discrete wavelet transform-enhanced deep separable sub-network at the same time, and perform feature extraction in the time domain, frequency domain and time-frequency domain respectively; Step S22: input the features obtained in step S21 into the dynamic multi-head attention mechanism module for interactive learning, capture the intrinsic connection and complementary information between different features, and fuse the features of interactive learning; Step S23: Flatten the fused features into a one-dimensional vector and then input them into the Softmax layer, thus completing the feature learning and classification process of the entire model and obtaining a trained parallel multi-fusion deep learning network model.

2. The bearing fault diagnosis method based on the parallel multi-fusion deep learning network model according to claim 1, It is characterized in that In step S2, during the entire training process of the model, the cross entropy loss function is used to calculate the loss of the predicted and true labels, and the back-propagation algorithm is used to update the weight parameters.

3. The bearing fault diagnosis method based on the parallel multi-fusion deep learning network model according to claim 1, It is characterized in that Step S1 specifically includes the following steps: Step S11, obtain sample data set I: obtain the vibration signals of rolling bearings at the same sampling frequency, different loads, and different speeds through public data sets, the sampling frequency is 25.6KHz, and the rolling bearings include four states: inner ring fault, outer ring fault, rolling element fault, and normal state; take 500 data points as a sample data, and assign sample data labels according to the rolling bearing state, and combine multiple sample data to obtain sample data set I, 70% of the sample data set I is divided out as a training data set, and the remaining 30% is used as a test data set.

4. The rolling bearing fault diagnosis method based on the parallel multi-fusion deep learning network model according to claim 3, Features: The step S1 specifically further comprises the following steps: Step S12, obtain sample data set II: obtain the vibration signal of the rolling bearing at the same sampling frequency and different speeds through the public data set, the sampling frequency is 20KHz, and the rolling bearing includes four states: inner ring fault, outer ring fault, rolling element fault and normal state; take 500 data points as a sample data, and assign sample data labels according to the rolling bearing state, and combine multiple sample data to obtain sample data set II, 70% of the sample data set II is divided out as a training data set, and the remaining 30% is used as a test data set.

5. The rolling bearing fault diagnosis method based on the parallel multi-fusion deep learning network model according to claim 1, Features: The bidirectional gated recurrent unit sub-network adaptively learns dynamic information based on the direct input original signal and performs time domain feature extraction; The Fast Fourier Transform-Convolutional Neural Network sub-network uses Fast Fourier Transform to transform the original signal into a frequency domain signal, which is then input into the Convolutional Neural Network to extract frequency domain features. The discrete wavelet transform-enhanced deep separable sub-network uses discrete wavelet transform to transform the one-dimensional time series into two-dimensional time-frequency domain image features, and inputs the time-frequency domain image into the enhanced deep separable sub-network to extract time-frequency domain features.

6. The rolling bearing fault diagnosis method based on the parallel multi-fusion deep learning network model according to claim 1, Features: The fast Fourier transform-convolutional neural network subnetwork includes a fast Fourier transform layer, three convolution layers, and three maximum pooling layers. The signals output by each convolution layer enter the maximum pooling layer for processing. After the signal is output by the third convolution layer, it is processed by the maximum pooling layer and enters the dynamic multi-head attention mechanism module. The discrete wavelet transform-enhanced deep separable subnetwork includes a discrete wavelet transform layer, three deep separable convolutional layers, and three maximum pooling layers. The signals output by each deep separable convolutional layer enter the maximum pooling layer for processing. After the signal is output by the third deep separable convolutional layer, it enters the dynamic multi-head attention mechanism module through the maximum pooling layer. An SRM high-pass filter is set between the discrete wavelet transform layer and the first depth-wise separable convolutional layer, which makes it easier for the network to learn the key information in the image and enhances the high-frequency information.

7. The rolling bearing fault diagnosis method based on the parallel multi-fusion deep learning network model according to claim 1, Features: The calculation process of the dynamic multi-head attention mechanism module in step S22 is as follows: first, the input Query (Q), Key (K) and Value (V) are mapped to a new vector space through linear transformation. The dimensions of these vectors are d model , the formula is as follows: Q′=W Q ·Q,K′=W K ·K,V'=W V ·V In the formula, Q represents the query vector of the input data, K represents the key vector of the input data, V represents the value vector of the input data, and W Q , W K , W V is the learned weight matrix used to convert the input Q, K, V into a new feature representation; Q′ represents the query vector of the new feature, K′ represents the key vector of the new feature, and V′ represents the value vector of the new feature; Introduce the dynamic adjustment parameter α to dynamically adjust Q and K. The formula is as follows: Where Q″ represents the query vector adjusted by the dynamic adjustment parameters, and K″ represents the key vector adjusted by the dynamic adjustment parameters; Then the transformed Q″, K″, V′ are divided into h heads, each head has its own Q, K, V, and the dimension of each head is d k , the formula is as follows: Q i =split_heads(Q”),K i =split_heads(K”),V i =split_heads(V') In the formula, Q i represents the query vector of the i-th head, K i represents the key vector of the i-th head, V i Represents the value vector of the i-th head, i represents the i-th head, 1≤i≤h, split_heads is the split function, and the dimension of each head is i.e. k ; Calculate the attention weight of each head and normalize it. The calculation formula is as follows: In the formula, α i represents the dynamic adjustment parameter of the i-th head; Calculate the input of each self-attention mechanism as follows: Attention i =Attention_Weight i ·V i The output of each head is concatenated together and linearly transformed to get the final output: Output=Concat(Attention 1 ,Attention 2 ,...,Attention h )·W O Where W O is the output linear transformation weight matrix.

8. The rolling bearing fault diagnosis method based on the parallel multi-fusion deep learning network model according to claim 1, Features: Step S3 specifically includes the following steps: Step S31, inputting the test data set in step S1 into the trained bidirectional gated recurrent unit sub-network, fast Fourier transform-convolutional neural network sub-network and discrete wavelet transform-enhanced deep separable sub-network at the same time to obtain a feature parameter matrix; Step S32: Input the feature parameter matrix obtained in step S31 into the trained dynamic multi-head attention mechanism module for interactive learning, capture the intrinsic connection and complementary information between different features, fuse the features of interactive learning, flatten the fused features into a one-dimensional vector, and then input it into the Softmax layer, and output the bearing state category of the identified test data set; Step S33: Compare the identified bearing status category of the test data set with the sample data label of the test data set, and calculate the diagnosis and recognition accuracy of the rolling bearing fault.

Images