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

Abstract

The application discloses a bearing fault diagnosis method based on a parallel multi-fusion deep learning network model, and comprises the following steps: obtaining a sample data set and dividing the sample data set into a training data set and a test data set, training the parallel multi-fusion deep learning network model by using the training data set, and inputting the test data set into the trained parallel multi-fusion deep learning network model to perform fault diagnosis. The parallel multi-fusion deep learning network model comprises three parallel sub-networks, namely, a bidirectional gate 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 three parallel sub-networks and the dynamic multi-head attention mechanism module are combined to automatically extract and identify fault features of rolling bearings under multiple working conditions, so that adaptive fault feature extraction under different domain signals and rolling bearing fault diagnosis under complex working conditions are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent bearing fault diagnosis technology, specifically a bearing fault diagnosis method based on a parallel multi-fusion deep learning network model. Background Technology

[0002] Rolling bearings, as key components in mechanical equipment, frequently operate under high-speed, heavy-load, high-temperature, and variable working environments, making them highly susceptible to failure. Statistics show that approximately 30% to 40% of mechanical failures are caused by bearings, resulting not only in safety hazards but also economic losses. Because vibration signals can reflect rich state information, vibration signal-based fault diagnosis technology has wide applications in bearing fault diagnosis. With the development of artificial intelligence technology, intelligent fault diagnosis methods have attracted considerable attention from researchers. Traditional intelligent fault diagnosis processes typically include the following steps: first, acquiring and preprocessing vibration signals; then, extracting features from the processed signals; and finally, using a machine learning-based classifier to identify the bearing's health status. However, feature extraction in this traditional method relies on expert experience and advanced signal processing techniques, making the selection of a suitable set of features very difficult, and machine learning cannot uncover deep-level 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 existing methods focus on combining a single feature extraction input with a single network, and these methods largely ignore the temporal characteristics of the original vibration signal. Only features from a single domain can be extracted, resulting in incomplete feature extraction and an inability to fully reflect fault characteristics. In actual engineering, due to the influence of operating conditions and environment, the collected vibration signals often exhibit different characteristics in different domains. Relying solely on features from a single domain may miss some information, thus affecting the accuracy of the final diagnosis. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this 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 method automatically extracts and identifies fault features of rolling bearings under multiple operating conditions, achieving adaptive fault feature extraction under different domain signals and fault diagnosis of rolling bearings under complex operating conditions.

[0004] To achieve the above objectives, the specific solution adopted by the present invention is as follows:

[0005] The bearing fault diagnosis method based on a parallel multi-fusion deep learning network model mainly includes the following steps:

[0006] Step S1: Obtain vibration signals of the bearing under multiple operating conditions through a public dataset, label the collected vibration signals and segment them to obtain a sample dataset of vibration signals, which includes a training dataset and a test dataset; Step S2: Train the parallel multi-fusion deep learning network model using the training dataset;

[0007] Step S3: Input the test dataset into the trained parallel multi-fusion deep learning network model to obtain the fault diagnosis results of the test data;

[0008] In step S2, the parallel multi-fusion deep learning network model includes three parallel sub-networks: 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, as well as a dynamic multi-head attention mechanism module. The specific method for training the parallel multi-fusion deep learning network model using the training dataset is as follows:

[0009] Step S21: Simultaneously input the training dataset into the bidirectional gated recurrent unit subnetwork, the fast Fourier transform-convolutional neural network subnetwork, and the discrete wavelet transform-enhanced deep separable subnetwork to 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 connections and complementary information between different features, and fuse the interactively learned features;

[0010] Step S23: Flatten the fused features into a one-dimensional vector and then input it into the Softmax layer. This completes the feature learning and classification process of the entire model, resulting in a trained parallel multi-fusion deep learning network model.

[0011] Furthermore, in step S2, throughout the entire training process of the model, the cross-entropy loss function is used to calculate the loss between the prediction and the true label, and the backpropagation algorithm is used to update the weight parameters.

[0012] Furthermore, step S1 specifically includes the following steps:

[0013] Step S11: Obtain Sample Dataset I: Obtain vibration signals of rolling bearings at the same sampling frequency, different loads, and different speeds through a public dataset. The sampling frequency is 25.6 kHz. The rolling bearings include four states: inner ring fault, outer ring fault, rolling element fault, and normal state. Each sample data point consists of 500 data points, and the sample data is labeled according to the rolling bearing state. Multiple sample data are combined to obtain Sample Dataset I. 70% of Sample Dataset I is used as the training dataset, and the remaining 30% is used as the test dataset.

[0014] Furthermore, step S1 specifically includes the following steps:

[0015] Step S12: Obtain Sample Dataset II: Obtain vibration signals of rolling bearings at the same sampling frequency but different rotational speeds through a public dataset. The sampling frequency is 20kHz. The rolling bearings include four states: inner ring fault, outer ring fault, rolling element fault, and normal state. Each sample data is composed of 500 data points, and the sample data is labeled according to the rolling bearing state. Multiple sample data are combined to obtain Sample Dataset II. 70% of Sample Dataset II is used as the training dataset, and the remaining 30% is used as the test dataset.

[0016] Furthermore, the bidirectional gated recurrent unit subnetwork adaptively learns dynamic information based on the directly input raw signal and performs temporal feature extraction;

[0017] The Fast Fourier Transform-Convolutional Neural Network subnetwork uses Fast Fourier Transform to transform the original signal into a frequency domain signal, and then inputs it into a convolutional neural network to extract frequency domain features.

[0018] Discrete wavelet transform-enhanced deep separable subnetwork uses discrete wavelet transform to transform a 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.

[0019] Furthermore, the Fast Fourier Transform-Convolutional Neural Network subnetwork includes a Fast Fourier Transform layer, three convolutional layers, and three max pooling layers; the signals output from each convolutional layer are processed by the max pooling layer, and after being output from the third convolutional layer, the signals are processed by the max pooling layer and then enter the dynamic multi-head attention mechanism module.

[0020] The Discrete Wavelet Transform-Enhanced Deeply Separable Subnetwork includes a Discrete Wavelet Transform layer, three Deeply Separable Convolutional Layers, and three Max Pooling Layers. The signals output from each Deeply Separable Convolutional Layer are processed by the Max Pooling Layer. After the signal is output from the third Deeply Separable Convolutional Layer, it is processed by the Max Pooling Layer and then enters the Dynamic Multi-Head Attention Mechanism Module.

[0021] An SRM high-pass filter is placed between the discrete wavelet transform layer and the first depth-separable convolutional layer, which makes it easier for the network to learn key information in the image and enhances high-frequency information.

[0022] Furthermore, the calculation process of the dynamic multi-head attention mechanism module in step S22 is as follows:

[0023] First, the input Query(Q), Key(K), and Value(V) are mapped to a new vector space through a linear transformation. These vectors all have a dimension of d. modelThe formula is as follows:

[0024] Q' = W Q ·Q,K'=W K ·K,V'=W V ·V

[0025] 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 represents the query vector of the input data. Q W K W V It is the learned weight matrix, used to transform 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;

[0026] Introducing a dynamic adjustment parameter α, Q and K are dynamically adjusted, as shown in the following formula:

[0027]

[0028] In the formula, 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″, and V′ are divided into h heads, each head having its own Q, K, and V, and each head has a dimension of d. k The formula is as follows:

[0029] Q i =split_heads(Q”),K i =split_heads(K”),V i =split_heads(V')

[0030] In the formula, Q i Let K represent the query vector for the i-th head. i V represents the key vector of the i-th head. i Let represent the value vector of the i-th head, where i represents the i-th head, 1 ≤ i ≤ h, and split_heads is the splitting function. The dimension of each head is . That is, d k ;

[0031] The attention weights for each head are calculated and normalized using the following formula:

[0032]

[0033] In the formula, α i This represents the dynamic adjustment parameter of the i-th head;

[0034] The input for each self-attention mechanism is calculated using the following formula:

[0035] Attention i =Attention_Weight i ·V i

[0036] The outputs of each head are concatenated together and subjected to a linear transformation to obtain the final output:

[0037] Output=Concat(Attention1,Attention2,...,Attention h )·W O

[0038] In the formula, W O It is the output linear transformation weight matrix.

[0039] Furthermore, step S3 specifically includes the following steps:

[0040] Step S31: Simultaneously input the test dataset from Step S1 into the trained bidirectional gated recurrent unit subnetwork, fast Fourier transform-convolutional neural network subnetwork, and discrete wavelet transform-enhanced deep separable subnetwork to obtain the 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 connections and complementary information between different features, and fuse the interactively learned features. 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 discriminated test dataset;

[0041] Step S33: Compare the bearing condition category of the identified test dataset with the sample data label of the test dataset to calculate the diagnostic accuracy of the rolling bearing fault.

[0042] Beneficial effects:

[0043] (1) The bearing fault diagnosis method disclosed in this 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 realize the automatic extraction and identification of rolling bearing fault features under different domain signals, which greatly improves the accuracy of rolling bearing fault diagnosis under complex working conditions.

[0044] (2) The bearing fault diagnosis method disclosed in this invention has a bidirectional gated cyclic unit sub-network, which can adaptively learn dynamic information.

[0045] (3) The bearing fault diagnosis method disclosed in this invention uses fast Fourier transform to convert the original signal into a frequency domain signal and discrete wavelet transform to convert a one-dimensional time series into two-dimensional time-frequency domain image features, which enriches the diversity of features extracted by the parallel multi-fusion deep learning network model.

[0046] (4) The bearing fault diagnosis method disclosed in this invention includes a Fast Fourier Transform-Convolutional Neural Network subnetwork comprising a Fast Fourier Transform layer, three convolutional layers, and three max pooling layers; and a Discrete Wavelet Transform-Enhanced Depth Separable subnetwork comprising a Discrete Wavelet Transform layer, three depth separable convolutional layers, and three max pooling layers; wherein, an SRM high-pass filter is provided between the Discrete Wavelet Transform layer and the first depth separable convolutional layer, which makes it easier for the network to learn key information in the image and enhances high-frequency information.

[0047] (5) The bearing fault diagnosis method disclosed in this invention has a dynamic multi-head attention mechanism module that can interact and fuse the features of the three extracted branches, capture the intrinsic connection and complementary information between different features, and enhance the feature expression ability. Attached Figure Description

[0048] Figure 1 This is a flowchart of the bearing fault diagnosis process of the present invention.

[0049] Figure 2 This is a structural diagram of the depth-separable convolutional layer of the present invention.

[0050] Figure 3 This is a structural diagram of the dynamic multi-head attention mechanism module of the present invention.

[0051] Figure 4 This is a structural diagram of the bearing fault diagnosis model of the present invention.

[0052] Figure 5 This is the confusion matrix diagram of sample dataset I.

[0053] Figure 6 This is a T-SNE visualization of sample dataset I.

[0054] Figure 7 This is the confusion matrix diagram of sample dataset II.

[0055] Figure 8 This is a T-SNE visualization of sample dataset II.

[0056] Figure 9 This is a comparison chart of the diagnostic accuracy of sample dataset I and the comparative model.

[0057] Figure 10This is a comparison chart of the diagnostic accuracy of sample dataset II and the comparative model. Detailed Implementation

[0058] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0059] Example 1

[0060] This embodiment discloses a rolling bearing fault diagnosis method based on a parallel multi-fusion deep learning network model. The following section combines... Figure 1-4 Each step is explained in detail.

[0061] Step S1: Obtain sample dataset: Obtain vibration signals of rolling bearings under multiple operating conditions through public datasets, label and segment the collected vibration signals to obtain a sample dataset of vibration signals, which includes training dataset and test dataset.

[0062] Step S1 includes the following steps:

[0063] Step S11: Obtain Sample Dataset I: Obtain vibration signals of rolling bearings at the same sampling frequency, different loads, and different speeds through public datasets (Huazhong University of Science and Technology Rolling Bearing Dataset and University of Cincinnati Rolling Bearing Dataset). The sampling frequency is 25.6 kHz. Rolling bearings include four states: inner ring fault, outer ring fault, rolling element fault, and normal state. In this embodiment, Sample Dataset I includes the operating results of the bearing under three different speed conditions, namely 65 Hz, 70 Hz, and 75 Hz. Each sample data is composed of 500 data points, and the sample data is labeled according to the rolling bearing state. Multiple sample data are combined to obtain Sample Dataset I. The specific details of Sample Dataset I in this embodiment are shown in Table 1.

[0064] Table 1. Details of Sample Dataset I

[0065]

[0066] Step S12: Obtain Sample Dataset II: Obtain vibration signals of rolling bearings at the same sampling frequency but different rotational speeds using a publicly available dataset. The sampling frequency is 20kHz. The rolling bearings are classified into four states: inner ring fault, outer ring fault, rolling element fault, and normal state. Each sample dataset consists of 500 data points, and labels are assigned to the sample data according to the rolling bearing state. Multiple sample datasets are combined to obtain Sample Dataset II. Details of Sample Dataset II are shown in Table 2.

[0067] Table 2 details the sample dataset II

[0068]

[0069]

[0070] Step S13: Divide 70% of sample dataset I and sample dataset II into training datasets and use the remaining 30% as test datasets.

[0071] Step S2: Establish a parallel multi-fusion deep learning network model and train the parallel multi-fusion deep learning network model using the training dataset.

[0072] Step S2 will be explained in detail below.

[0073] Step S21: Establish a parallel multi-fusion deep learning network model. This model includes three parallel sub-networks: 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; a dynamic multi-head attention mechanism module that aggregates the output signals of the three parallel sub-networks; a fully connected layer; and a softmax layer. Specifically, the fast Fourier transform-convolutional neural network sub-network sequentially includes a fast Fourier transform layer, convolutional layer 1, max pooling layer, convolutional layer 2, max pooling layer, convolutional layer 3, and max pooling layer; the discrete wavelet transform-enhanced deep separable sub-network sequentially includes a discrete wavelet transform layer, an SRM filter, a deep separable convolutional layer 1, a max pooling layer, a deep separable convolutional layer 2, a max pooling layer, a deep separable convolutional layer 3, and a max pooling layer.

[0074] Step S22: Simultaneously input the training dataset into the bidirectional gated recurrent unit subnetwork, the fast Fourier transform-convolutional neural network subnetwork, and the discrete wavelet transform-enhanced deep separable subnetwork to extract features in the time domain, frequency domain, and time-frequency domain. Specifically, the bidirectional gated recurrent unit subnetwork adaptively learns dynamic information based on the directly input raw signal to extract time-domain features; the fast Fourier transform-convolutional neural network subnetwork uses fast Fourier transform to convert the raw signal into a frequency domain signal before inputting 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 a 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, making it easier for the network to learn key information in the image and enhancing high-frequency information. The SRM high-pass filter is:

[0075] 1st:[-1 1],2nd: 3rd:

[0076] edge 3×3 : square 3×3 :

[0077] edge 5×5 : square 5×5 :

[0078] The seven high-pass filters described above can extract high-frequency characteristic information from seven different modes. All seven high-pass filters are then expanded to a size of 5×5.

[0079] Step S23: Input the features obtained in step S22 into the dynamic multi-head attention mechanism module for interactive learning, capture the intrinsic connections and complementary information between different features, and fuse the interactively learned features. The scaling factor of traditional multi-head attention mechanisms is fixed, making the model lack flexibility. Therefore, a parameter α is introduced. When calculating attention weights, 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:

[0080] First, the input Query(Q), Key(K), and Value(V) are mapped to a new vector space through a linear transformation. These vectors all have a dimension of d. model The formula is as follows:

[0081] Q' = W Q ·Q,K'=W K ·K,V'=W V ·V

[0082] 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 represents the query vector of the input data. Q W K W V It is the learned weight matrix used to transform 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.

[0083] Next, a dynamic adjustment parameter α is introduced to dynamically adjust Q and K, as shown in the following formula:

[0084]

[0085] In the formula, Q″ represents the query vector after being adjusted by the dynamic adjustment parameters, and K″ represents the key vector after being adjusted by the dynamic adjustment parameters.

[0086] The transformed Q″, K″, V′ are divided into h heads, each head having its own Q, K, V, and each head having a dimension of d. k The formula is as follows:

[0087] Q i =split_heads(Q”),K i =split_heads(K”),V i =split_heads(V')

[0088] In the formula, Q i Let K represent the query vector for the i-th head. i V represents the key vector of the i-th head. i Let represent the value vector of the i-th head, where i represents the i-th head, 1 ≤ i ≤ h, and split_heads is the splitting function. The dimension of each head is . That is, d k .

[0089] The attention weights for each head are calculated and normalized using the following formula:

[0090]

[0091] In the formula, α i This represents the dynamic adjustment parameter of the i-th head;

[0092] The input for each self-attention mechanism is calculated using the following formula:

[0093] Attention i =Attention_Weight i ·V i

[0094] The outputs of each head are concatenated together and subjected to a linear transformation to obtain the final output:

[0095] Output=Concat(Attention1,Attention2,...,Attention h )·W O

[0096] In the formula, W O It is the output linear transformation weight matrix.

[0097] Step S24: Flatten the fused features into a one-dimensional vector and then input it into the Softmax layer to complete the feature learning and classification process of the entire model.

[0098] It should be noted that during the entire training process, the cross-entropy loss function is used to calculate the loss between the prediction and the true label, and the backpropagation algorithm is used to update the weight parameters.

[0099] The iteration count is 50 times, the batch size is 32, and the parameters of the bidirectional gated recurrent unit subnetwork, the fast Fourier transform-convolutional neural network subnetwork, the discrete wavelet transform-enhanced deep separable subnetwork, and the dynamic multi-head attention mechanism module are updated iteratively using the backpropagation algorithm until the iteration count is reached. This indicates that the parallel multi-fusion deep learning network model training is complete, and the process proceeds to step S3.

[0100] Step S3: Input the test dataset into the trained parallel multi-fusion deep learning network model to obtain the fault diagnosis results of the test data.

[0101] 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 in step S31, the trained bidirectional gated recurrent unit subnetwork, fast Fourier transform-convolutional neural network subnetwork, and discrete wavelet transform-enhanced deep separable subnetwork are obtained. The test dataset described in S1 is then input into the trained bidirectional gated recurrent unit subnetwork, fast Fourier transform-convolutional neural network subnetwork, and discrete wavelet transform-enhanced deep separable subnetwork to obtain the feature parameter matrix.

[0102] 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 relationship and complementary information between different features, fuse the interactively learned features, 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 discriminated test dataset.

[0103] Step S33: Compare the bearing condition category of the identified test dataset with the sample data label of the test dataset to calculate the diagnostic accuracy of the rolling bearing fault.

[0104] The confusion matrices and T-SNE visualizations of sample datasets I and II are shown below. Figures 5 to 8 As shown. Figure 5 In the confusion matrix, 0-11 correspond to 12 state categories. The values ​​on the diagonal represent the recognition accuracy for each state, while the remaining values ​​represent the recognition error rate. The results of the confusion matrix show that, except for category 7 which had a 3.33% misdiagnosis as category 4, the diagnosis rate for all other categories was 100%, with no misdiagnoses. Figure 7The confusion matrix results showed that the diagnostic rate for all categories was 100%, with no misdiagnoses. Figure 6 and Figure 8 The T-SNE feature visualization showed clear boundaries between categories. Even after reducing the dimensionality from high-dimensional features to a two-dimensional view, each fault category remained highly distinguishable. This verifies that the proposed method exhibits extremely high classification accuracy and stability in the field of rolling bearing fault identification, as well as the ability to effectively extract and preserve classification differences in complex feature spaces, effectively reducing confusion between categories.

[0105] To verify the performance of the proposed method, it was compared with traditional CNN, MSCNN, AlexNet, and advanced TSCNN and MSTCNN network models. Specifically, (1) CNN, MSCNN, and AlexNet used the original vibration signal as direct input; (2) TSCNN was a parallel network model with inputs from two different domains: time-domain and time-frequency domain; and (3) MSTCNN was a parallel network model with inputs from three different domains: time-domain, frequency domain, and time-frequency domain. To ensure fairness, all models were trained using the same initial parameters as the proposed method. The evaluation criteria were accuracy, F1 score, precision, and recall. To reduce the influence of randomness, each trial was repeated three times.

[0106] First, dataset I was used for validation, and the diagnostic accuracy of each model was obtained as shown in Table 3. Figure 9 As shown.

[0107] Table 3 shows the diagnostic rates of each model when using dataset I.

[0108] Model Average F1 score / % Average accuracy / % Average recall rate / % 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 This invention model 99.63 99.64 99.63

[0109] From Table 3 and Figure 9 As can be seen, the model proposed in this invention demonstrates superior performance compared to the comparison models across all four evaluation metrics (average F1 score, average precision, average recall, and average accuracy). Among the six models compared, only the MSTCNN model failed to break the 50% threshold, while the other models achieved over 70% in all metrics except for the average F1 score of TSCNN. Compared to the other five models, the model proposed in this invention achieves a significant improvement in average diagnostic accuracy, specifically: compared to the comparison models, the proposed model achieves performance improvements of 20.37%, 13.15%, 1.02%, 24.54%, and 50% in average diagnostic accuracy, respectively. This result proves that the method proposed in this invention not only performs excellently in effectiveness but also has significant advantages in generalization ability and robustness.

[0110] Then, dataset II was used for further validation, and the diagnostic rates of each model are shown in Table 4. Figure 10 As shown.

[0111] Table 4 shows the diagnostic rates of each model when using Dataset II.

[0112] Model Average F1 score / % Average accuracy / % Average recall rate / % 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 This invention model 100 100 100

[0113] From Table 4 and Figure 10 As can be seen, among the six models compared, MSTCNN achieved over 80% accuracy in all metrics except for average precision. The model proposed in this invention significantly improves average diagnostic accuracy compared to the other five models, specifically achieving performance improvements of 5.46%, 9.84%, 4.56%, 5.10%, and 15.30% in average diagnostic accuracy, respectively, compared to the comparison models. This result demonstrates that the method proposed in this invention not only performs excellently in effectiveness but also exhibits significant advantages in generalization ability and robustness. Furthermore... Figure 10 It can be clearly seen that the stability of the method proposed in this invention is significantly improved.

[0114] To verify the effectiveness of each branch in the proposed method, ablation experiments were conducted. The comparative analysis models 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 network parameters were set the same as those in the proposed method. The results for dataset I are shown in Table 5.

[0115] Table 5 shows the diagnostic rates of each model when using dataset I.

[0116]

[0117]

[0118] As shown in Table 5, the results of all four evaluation metrics of the proposed method exceed 99%. Removing any one branch significantly reduces the accuracy of the results. This is because the proposed method can extract multiple features of the vibration signal, such as time, frequency, and local features, improving the robustness and generalization of the model. This result proves the indispensability of each branch in the proposed method. Moreover, the results are significantly reduced when all branches are used individually. In particular, the time-frequency domain branch does not exceed 65%, demonstrating the synergistic interaction between the various branches.

[0119] Then, dataset II was used for further validation, and the diagnostic accuracy is shown in Table 6.

[0120] Table 6 shows the diagnostic rates of each model when using Dataset II.

[0121] Model Average accuracy / % Average F1 score / % Average accuracy / % Average recall rate / % 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 This invention model 100 100 100 100

[0122] As shown in Table 6, the proposed method achieved 100% accuracy across all four evaluation metrics. Removing any one branch, while maintaining accuracy above 90%, significantly reduced it compared to the proposed method. This demonstrates the indispensability of each branch in the proposed method. Furthermore, the results also significantly decreased when all branches were used individually.

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention in any way. All equivalent transformations or modifications made in accordance with the essence of the present invention should be covered within the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method based on a parallel multi-fusion deep learning network model, characterized in that, The main steps include the following: Step S1: Obtain vibration signals of the bearing under multiple operating conditions through a public dataset, label the collected vibration signals and segment them to obtain a sample dataset of vibration signals, which includes a training dataset and a test dataset. Step S2: Train the parallel multi-fusion deep learning network model using the training dataset; Step S3: Input the test dataset into the trained parallel multi-fusion deep learning network model to obtain the fault diagnosis results of the test data; In step S2, the parallel multi-fusion deep learning network model includes three parallel sub-networks: 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, as well as a dynamic multi-head attention mechanism module. The specific method for training the parallel multi-fusion deep learning network model using the training dataset is as follows: Step S21: Simultaneously input the training dataset into the bidirectional gated recurrent unit subnetwork, the fast Fourier transform-convolutional neural network subnetwork, and the discrete wavelet transform-enhanced deep separable subnetwork, 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 connections and complementary information between different features, and fuse the interactively learned features; Step S23: Flatten the fused features into a one-dimensional vector and then input it into the Softmax layer. This completes the feature learning and classification process of the entire model, resulting in a trained parallel multi-fusion deep learning network model.

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

3. The bearing fault diagnosis method based on a parallel multi-fusion deep learning network model according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S11: Obtain Sample Dataset I: Obtain vibration signals of rolling bearings at the same sampling frequency, different loads, and different speeds through a public dataset. The sampling frequency is 25.6 kHz. The rolling bearings include four states: inner ring fault, outer ring fault, rolling element fault, and normal state. Each sample data is composed of 500 data points, and the sample data is labeled according to the rolling bearing state. Multiple sample data are combined to obtain Sample Dataset I. 70% of Sample Dataset I is used as the training dataset, and the remaining 30% is used as the test dataset.

4. The bearing fault diagnosis method based on a parallel multi-fusion deep learning network model according to claim 3, characterized in that: Step S1 further includes the following steps: Step S12: Obtain Sample Dataset II: Obtain vibration signals of rolling bearings at the same sampling frequency but different rotational speeds through a public dataset. The sampling frequency is 20kHz. The rolling bearings include four states: inner ring fault, outer ring fault, rolling element fault, and normal state. Each sample data is composed of 500 data points, and the sample data is labeled according to the rolling bearing state. Multiple sample data are combined to obtain Sample Dataset II. 70% of Sample Dataset II is used as the training dataset, and the remaining 30% is used as the test dataset.

5. The bearing fault diagnosis method based on a parallel multi-fusion deep learning network model according to claim 1, characterized in that: The bidirectional gated recurrent unit subnetwork adaptively learns dynamic information based on the directly input raw signal and performs temporal feature extraction. The Fast Fourier Transform-Convolutional Neural Network subnetwork uses Fast Fourier Transform to transform the original signal into a frequency domain signal, and then inputs it into a convolutional neural network to extract frequency domain features. Discrete wavelet transform-enhanced deep separable subnetwork uses discrete wavelet transform to transform a 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.

6. The bearing fault diagnosis method based on a parallel multi-fusion deep learning network model according to claim 1, characterized in that: The Fast Fourier Transform-Convolutional Neural Network subnetwork includes a Fast Fourier Transform layer, three convolutional layers, and three max pooling layers. The signals output from each convolutional layer are processed by the max pooling layer. After the signal is output from the third convolutional layer, it is processed by the max pooling layer and then enters the dynamic multi-head attention mechanism module. The Discrete Wavelet Transform-Enhanced Deeply Separable Subnetwork includes a Discrete Wavelet Transform layer, three Deeply Separable Convolutional Layers, and three Max Pooling Layers. The signals output from each Deeply Separable Convolutional Layer are processed by the Max Pooling Layer. After the signal is output from the third Deeply Separable Convolutional Layer, it is processed by the Max Pooling Layer and then enters the Dynamic Multi-Head Attention Mechanism Module. An SRM high-pass filter is placed between the discrete wavelet transform layer and the first depth-separable convolutional layer, which makes it easier for the network to learn key information in the image and enhances high-frequency information.

7. The bearing fault diagnosis method based on a parallel multi-fusion deep learning network model according to claim 1, characterized in that: 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 a linear transformation. The dimensions of these vectors are all... d model The formula is as follows: In the formula, Q This represents the query vector of the input data. K This represents the key vector of the input data. V W represents the value vector of the input data. Q W K W V It is the learned weight matrix, used to weight the input... Q , K , V Convert to a new feature representation; Q′ A query vector representing a new feature. K′ The key vector representing the new feature. V′ A value vector representing the new feature; Introducing a dynamically adjustable parameter α, for Q and K The formula for dynamic adjustment is as follows: In the formula, Q″ This represents the query vector after dynamically adjusting parameters. K″ This represents the key vector after dynamic parameter adjustment; Next, the transformed Q″, K″, V′ are divided into h heads, each head having its own... Q , K , V The dimensions of each head are d k The formula is as follows: In the formula, Q i Indicates the first i The query vector of the size, K i Indicates the first i The key vector of the head, V i Indicates the first i A vector of values ​​for each head. i Indicates the first i Height, 1≤ i ≤ h `split_heads` is the splitting function, and the dimension of each head is... ,Right now d k ; The attention weights for each head are calculated and normalized using the following formula: In the formula, α i Indicates the first i Dynamic adjustment parameters for size; The input for each self-attention mechanism is calculated using the following formula: The outputs of each head are concatenated together and subjected to a linear transformation to obtain the final output: In the formula, W O It is the output linear transformation weight matrix.

8. The bearing fault diagnosis method based on a parallel multi-fusion deep learning network model according to claim 1, characterized in that: Step S3 specifically includes the following steps: Step S31: Input the test dataset from step S1 into the trained bidirectional gated recurrent unit subnetwork, fast Fourier transform-convolutional neural network subnetwork, and discrete wavelet transform-enhanced deep separable subnetwork simultaneously to obtain the 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 relationship and complementary information between different features, fuse the interactive learning features, 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 discriminated test dataset. Step S33: Compare the bearing condition category of the identified test dataset with the sample data label of the test dataset to calculate the diagnostic accuracy of the rolling bearing fault.

Images