Lightweight bearing fault diagnosis method based on multi-scale feature fusion

Abstract

A lightweight bearing fault diagnosis method based on multi-scale feature fusion is disclosed. The method involves collecting bearing vibration signals and inputting them into a trained fault diagnosis model for fault diagnosis. The fault diagnosis model comprises a multi-layer feature extraction module and an output layer. The feature extraction module employs a residual structure. In the main branch of the residual structure, the input data is first subjected to point-by-point convolution. The resulting data is then split into two paths, one of which is retained while the other is subjected to channel-by-channel atrous convolution and then concatenated with the retained data from the point-by-point convolution step. A channel-attention mechanism is introduced in the secondary branch of the residual structure to process the input data, which is then fused with the processed data from the main branch as the output of the feature extraction module. By utilizing the complementary nature of the channel-by-channel atrous convolution and point-by-point convolution, the method can filter out noise in the holes, capture long-range induced correlation features of the signal, and capture differential features between different channels, thereby improving fault diagnosis accuracy.

Description

Technical Field

[0001] The present invention relates to a bearing fault diagnosis method, and in particular to a lightweight bearing fault diagnosis method based on multi-scale feature fusion. Background Art

[0002] Rolling bearings, core components of industrial machinery, are widely used in rotating equipment such as motors, engines, and gearboxes. However, rolling bearings are prone to failure due to long-term operation under complex conditions such as high speeds and heavy loads. According to statistics, 45%-55% of mechanical equipment failures are related to bearings. Rolling bearing failures can cause property damage at the very least, or even casualties at the worst. Therefore, developing a highly accurate and feasible bearing fault diagnosis method is extremely important.

[0003] Deep learning, owing to its adaptive feature learning capabilities, is currently widely used in mechanical fault diagnosis. In real-world industrial environments, mechanical equipment often operates under complex circumstances. The signals collected by sensors are inevitably contaminated with noise and nonlinear interference. To improve the model's noise immunity and diagnostic accuracy, existing technologies employ multi-head attention, designing an attention module that fully considers the characteristics of rolling bearing faults, enhancing fault-related features and ignoring irrelevant ones. However, the inclusion of multi-head attention results in the model's FLOPs reaching tens of millions. Other approaches employ bearing fault feature extraction networks based on dynamic convolution and depthwise separable convolution, while also using a spatial attention mechanism (SAM) to adjust the weights of the output feature maps, allowing the model to focus more on useful information. However, the modules proposed by these approaches are too complex and costly to be deployed in practice.

[0004] Although the methods used in existing technologies have achieved certain results, they still have shortcomings in balancing recognition accuracy and model lightweightness. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to improve the anti-noise interference ability of the model and provide a lightweight bearing fault diagnosis method based on multi-scale feature fusion.

[0006] The technical solution adopted by the present invention to solve the above technical problems is: a lightweight bearing fault diagnosis method based on multi-scale feature fusion, collecting vibration signals during the operation of the bearing, and inputting them into the trained fault diagnosis model for fault diagnosis. The fault diagnosis model includes a multi-layer feature extraction module and an output layer that outputs the fault type according to the extracted features. The feature extraction module adopts a residual structure. The main branch of the residual structure first performs point-by-point convolution processing on the input data. The data after point-by-point convolution processing is divided into two paths, one of which is retained, and the other is processed by channel-by-channel hole convolution and then spliced ​​with the retained one path of data processed by point-by-point convolution in the channel dimension; the channel attention mechanism is introduced in the secondary branch of the residual structure to process the input data, and then it is fused with the data processed by the main branch as the output of the feature extraction module.

[0007] Furthermore, a wide kernel convolution layer is used to perform data compression and channel expansion on the collected original vibration signal, and then input it into the feature extraction module for feature extraction.

[0008] The convolution kernel size of the wide kernel convolution layer is 16, and the number of convolution kernel channels is 128.

[0009] The channel-by-channel dilated convolution of the feature extraction module adopts large-kernel channel-by-channel dilated convolution, and the convolution kernel size is 1×63.

[0010] Furthermore, the channel-by-channel atrous convolutions in the multi-layer feature extraction module have different dilation rates.

[0011] The fault diagnosis model includes four layers of sequentially connected feature extraction modules, and the expansion rates of the channel-by-channel dilated convolutions in the four layers of feature extraction modules are set to 1, 2, 5 and 1 respectively.

[0012] In the sub-branch of the residual structure of the feature extraction module, ECA attention is used to calculate the channel weights to obtain the importance of the input data, and then weights are assigned to each channel to obtain weighted data.

[0013] In the feature extraction module, the data after point-by-point convolution is processed using the PReLU activation function, and after batch normalization by the BN layer, it is divided into two paths for subsequent processing.

[0014] In the feature extraction module, the data after channel-by-channel hole convolution is processed using the PReLU activation function, and after batch normalization by the BN layer, it is spliced ​​with the retained data in the channel dimension.

[0015] The output layer includes a global average pooling layer, a fully connected layer and a Softmax layer.

[0016] The beneficial effects of the present invention are as follows: in the feature extraction module of the fault diagnosis model, feature extraction at different spatial positions, different distances, and different channels is achieved through point-by-point convolution, channel-by-channel dilated convolution, and a channel attention mechanism, and fully integrated, thereby more efficiently learning features and improving the model's ability to resist noise interference. By fusing the results of point-by-point convolution with the results of channel-by-channel dilated convolution, and utilizing the complementarity of channel-by-channel dilated convolution and point-by-point convolution, it is possible to filter out noise in the dilated areas, capture the long-range induced correlation features of the signal, and the differential features between different channels, thereby improving the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a structural diagram of the fault diagnosis model of the present invention.

[0018] Figure 2 It is a structural diagram of the feature extraction module of the present invention.

[0019] Figure 3 It is a line graph of recognition accuracy of different methods with different signal-to-noise ratios.

[0020] Figure 4 It is a t-SNE dimensionality reduction graph of the feature graph of the multi-layer feature extraction module of the present invention. DETAILED DESCRIPTION

[0021] The following is a clear and complete description of the technical solution of the present invention in conjunction with the accompanying drawings and specific embodiments. The specific contents listed in the following embodiments are not limited to the technical features required to solve the technical problems to be solved by the technical solutions described in the claims. At the same time, the examples listed are only part of the present invention, not all embodiments.

[0022] The lightweight bearing fault diagnosis method based on multi-scale feature fusion in the present invention relies on a trained fault diagnosis model. The input data of the fault diagnosis model is the vibration signal collected during the bearing movement process. The vibration signal is feature extracted and processed by the fault diagnosis model, and the fault type is output through its output layer.

[0023] The fault diagnosis model of the present invention includes multiple layers of feature extraction modules for extracting features from input data to learn fault features in depth. The structure of the feature extraction module is as follows: Figure 2As shown in the figure, a residual structure is used, which is a lightweight multi-scale dilated residual module, named MSL Block. The residual structure of the feature extraction module has two branches. The main branch includes a point-by-point convolution layer (Conv(1×1)) and a channel-by-channel dilated convolution layer. In the main branch, data of dimension C×W is first processed by a point-by-point convolution layer (Conv(1×1)) to learn the instantaneous features of the signal at close range and features at different spatial locations. The dimension is reduced to C / 2×W. The data is then split into two paths. One path is retained, while the other is processed by a channel-by-channel dilated convolution layer. This operation simultaneously performs dilated and channel-by-channel convolution, learning differential features across different channels and long-range correlation features. This increases the receptive field of the network feature extraction and fully captures the temporal correlation information of the signal. The data processed by the channel-by-channel dilated convolution is then concatenated with the previously retained data processed by the point-by-point convolution in the channel dimension to achieve multi-angle feature fusion. In the secondary branch, a channel attention mechanism is used to process input data. Specifically, ECA attention is used to calculate channel weights, determine the importance of input data, and enhance the model's sensitivity to important features. Weights are then assigned to each channel to generate weighted data. After the calculations of the main and secondary branches are completed, the fault feature information extracted by the two branches is further integrated using residuals. This not only alleviates network degradation but also enriches the fault feature information and improves the model's ability to extract fault features from multiple angles.

[0024] The channel-by-channel dilated convolution uses a large kernel channel-by-channel dilated convolution, and the convolution kernel size can be 1×63. Large kernel channel-by-channel convolution significantly increases the receptive field of the convolution layer, and retains the phase information of the signal, avoiding the loss of spatial position features in deep downsampling. Combining dilated convolution with large kernel channel-by-channel convolution can further improve the receptive field of the convolution layer. At the same time, the role of large kernel channel-by-channel dilated convolution is complementary to point-by-point convolution, which can filter out noise in the holes and capture the long-range induced correlation characteristics of the signal as well as the difference characteristics between different channels.

[0025] Figure 1In the illustrated embodiment, the fault diagnosis model utilizes a four-layer feature extraction module (MSL Block). This model uses a wide-kernel convolution layer to perform data compression and channel expansion on the collected raw vibration signal. This helps better capture low-frequency components in the signal, reduces the impact of high-frequency noise on fault diagnosis, and lays a solid foundation for subsequent feature learning. The wide-kernel convolution layer has a convolution kernel size of 16 and 128 channels. Wide-kernel convolution effectively compresses the input signal through a larger receptive field, reducing data redundancy. Channel expansion through wide-kernel convolution increases the number of channels in the feature map, enabling the network to learn more diverse features. The expanded feature map captures features at different frequencies and scales, enhancing the model's expressive power. Therefore, using wide-kernel convolution layers to perform data compression and channel expansion on the raw vibration signal not only effectively reduces interference from high-frequency noise but also increases feature diversity through channel expansion. This processing approach improves the model's noise immunity while reducing computational complexity, making the model more lightweight and efficient. Data processed by the wide-kernel convolution layer is then input into the feature extraction module for feature extraction. In this embodiment, four layers of feature extraction modules are stacked to process the data in sequence. The channel-by-channel dilated convolutions in different feature extraction modules can have different dilation rates, for example Figure 1 In the illustrated embodiment, the expansion rate of the first MSL Block is 1, the expansion rate of the second MSL Block is set to 2, the expansion rate of the third MSL Block is 5, and the expansion rate of the fourth MSL Block is 1. By setting different expansion rates at different layers, the grid effect caused by stacking multiple layers of dilated convolution can be avoided, effectively alleviating information loss and improving network performance.

[0026] Figure 1 Figure 2 shows the data processing process for the second MSL block, which is similar for other MSL blocks. In the figure, the data input to the second MSL block is of dimension C×W. After point-by-point convolution with kernel size k=1 in the main branch, the C×W data is reduced to C / 2×W. This data is retained while undergoing channel-by-channel atrous convolution with kernel size k=63. The result is concatenated with the retained data to form a C×W data set. The original input data to the MSL block is processed by the ECA attention mechanism in the secondary branch and then added to the processed data from the main branch as the output of the MSL block.

[0027] like Figure 2As shown in the figure, in the feature extraction module, the data after point-by-point convolution is processed using the PReLU activation function, and after batch normalization through the BN layer, it is divided into two paths for subsequent processing. The data after channel-by-channel dilated convolution is also processed using the PReLU activation function, and after batch normalization through the BN layer, it is concatenated with the retained data path in the channel dimension. The PReLU activation function is used to improve the nonlinear fitting ability of the model, and the BN layer is used to batch normalize the input of each layer to prevent gradient vanishing.

[0028] like Figure 1 As shown in the figure, the output layer of the fault diagnosis model includes a global average pooling layer (GAP), a fully connected layer (Linear), and a SoftMax layer. After feature extraction by the multi-layer feature extraction module, the learned features are globally average pooled, classified using the fully connected layer and the Softmax layer, and finally the fault type is output.

[0029] The fault diagnosis model was trained and validated using the CWRU bearing dataset. This dataset is categorized into 10 fault types based on fault location and damage severity. Four datasets, A, B, C, and D, are classified based on load. Datasets A, B, and C are for loads of 1HP, 2HP, and 3HP, respectively. Dataset D is a mixed dataset of loads 0-3. The vibration signals in this dataset are randomly overlapped and sampled with a step size of 207 and a sample length of 1024. Each dataset contains 500 single-class samples, for a total of 5000 samples. The training, validation, and test sets have a ratio of 8:1:1.

[0030] The experimental data parameter statistics are shown in Table 1 below:

[0031]

[0032] The following uses the above Figure 1 The fault diagnosis model of the embodiment is used to conduct a fault diagnosis experiment to test the accuracy of the fault diagnosis.

[0033] First, we conducted an ablation experiment, using the fault diagnosis model of the present invention as the baseline model. Six other comparison scenarios were set: no dilated convolution, no point-by-point convolution, no large-kernel channel-by-channel dilated convolution, no ECA attention mechanism (retaining residual connections), no ECA attention mechanism and residual connections, and using ReLU instead of PReLU. The experiment was conducted using Dataset D, and the experimental results are shown in Table 2 below:

[0034]

[0035] Table 2 shows that regardless of which module is missing from the fault diagnosis model, performance will decline, and the performance degradation of the comparison model becomes more severe as noise increases. The table also shows that when large-kernel per-channel dilated convolution is not used, the model's performance declines even more significantly. This demonstrates that long-range correlation features are crucial for bearing fault diagnosis based on vibration signals. The large receptive field provided by large-kernel per-channel dilated convolution is very helpful for long-range feature extraction. Dilated convolution can also effectively filter noise, resulting in improved performance.

[0036] To verify the noise immunity and complex working condition diagnostic capabilities of the proposed fault diagnosis model, we introduced WDCNN, MA1DCNN, DRSN-CW, and MIXCNN2 for comparison. By adding Gaussian white noise with varying signal-to-noise ratios to the CWRU datasets A, B, C, and D, we simulated the noise interference found in actual working environments and explored the fault recognition accuracy of different models in high-noise environments. This also further demonstrated the feature extraction capabilities of the proposed method. The specific results are shown in Table 3 below:

[0037]

[0038] In order to more intuitively show the accuracy of different models, a line graph of the fault recognition accuracy of different models for the CWRU dataset D under different noise environments is drawn, as shown in the figure below: Figure 3 shown.

[0039] From Table 3 above and Figure 3 It can be seen that compared with several other methods, the present invention can show better fault identification performance when facing the complex working condition data set D (mixed load). At the same time, when facing data sets A, B, and C that are not affected by load changes, the recognition accuracy of the present invention is higher.

[0040] In order to better understand the training process and feature extraction capability of the fault diagnosis model of the present invention, the t-SNE dimensionality reduction algorithm is used to reduce the dimension of the output feature map of the middle layer of the model. The results are as follows: Figure 4 As shown in the figure, after the first feature extraction module (MSL Block1), the ten categories tend to separate. After the second and third feature extraction modules (MSL Block2 and MSL Block3), there are obvious boundary areas between different categories, but the observation points in the same category are still relatively scattered. After the fourth feature extraction module (MSL Block4), the different categories are completely separated, and the observation points in the same category are closely clustered. This shows that the fault diagnosis model of the present invention is very effective in extracting features from raw data and has good anti-noise interference capabilities.

[0041] The following analysis focuses on model complexity. By comparing the number of parameters and FLOPs with other bearing fault diagnosis models or lightweight models, the lightweightness of the proposed fault diagnosis model is demonstrated. The number of parameters reflects the model's memory usage, while FLOPs reflects the model's computational complexity. The specific comparison results are shown in Table 4 below:

[0042]

[0043] As can be seen from Table 4 and Table 3, although the number of parameters and FLOPs of WDCNN are lower than those of the present invention, its accuracy is poor. MA1DCNN and DRSN-CW did not consider lightweighting during their design, and due to the existence of some special modules in their network structure, the computational complexity of the model is also very high. MIXCNN2 is a bearing fault diagnosis model with mixed information. This model uses depthwise separable convolution to reduce the number of parameters and has good accuracy in noisy environments. The parameter number of the present invention is 33% less than that of MIXCNN2, and the FLOPs are 27% less than that of MIXCNN2. In addition, the accuracy and noise resistance of the present invention are higher than those of MIXCNN2.

[0044] The above description of the specific embodiments is only used to help understand the technical concept and core ideas of the present invention. Although specific preferred embodiments are used herein to describe and illustrate the technical solutions, they should not be understood as limiting the present invention itself. Those skilled in the art may make various changes in form and details without departing from the technical concept of the present invention. These easily conceived changes or substitutions should all be included within the scope of protection of the present invention.

Claims

1. A lightweight bearing fault diagnosis method based on multi-scale feature fusion collects vibration signals during bearing operation and inputs them into a trained fault diagnosis model for fault diagnosis. The fault diagnosis model includes a multi-layer feature extraction module and an output layer that outputs the fault type based on the extracted features. The method is characterized by: The feature extraction module adopts a residual structure. In the main branch of the residual structure, the input data is first subjected to point-by-point convolution processing. The data after point-by-point convolution processing is divided into two paths, one of which is retained, and the other path is subjected to channel-by-channel dilated convolution processing and then spliced ​​with the retained one path that has undergone point-by-point convolution processing in the channel dimension; the channel attention mechanism is introduced into the secondary branch of the residual structure to process the input data, and then it is fused with the data processed by the main branch as the output of the feature extraction module; The channel-by-channel atrous convolutions in different feature extraction modules have different dilation rates.

2. The lightweight bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that: The wide kernel convolution layer is used to perform data compression and channel expansion on the collected original vibration signal, and then input it into the feature extraction module for feature extraction.

3. The lightweight bearing fault diagnosis method based on multi-scale feature fusion according to claim 2, characterized in that: The convolution kernel size of the wide kernel convolution layer is 16, and the number of convolution kernel channels is 128.

4. The lightweight bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that: The channel-by-channel dilated convolution of the feature extraction module adopts large-kernel channel-by-channel dilated convolution, and the convolution kernel size is 1×63.

5. The lightweight bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that: The fault diagnosis model includes four layers of sequentially connected feature extraction modules, and the expansion rates of the channel-by-channel dilated convolutions in the four layers of feature extraction modules are set to 1, 2, 5 and 1 respectively.

6. The lightweight bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that: In the sub-branch of the residual structure of the feature extraction module, ECA attention is used to calculate the channel weights to obtain the importance of the input data, and then weights are assigned to each channel to obtain weighted data.

7. The lightweight bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that: In the feature extraction module, the data after point-by-point convolution is processed using the PReLU activation function, and after batch normalization by the BN layer, it is divided into two paths for subsequent processing.

8. The lightweight bearing fault diagnosis method based on multi-scale feature fusion according to claim 7, characterized in that: In the feature extraction module, the data after channel-by-channel dilated convolution is processed using the PReLU activation function, and after batch normalization by the BN layer, it is spliced ​​with the retained data in the channel dimension.

9. The lightweight bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that: The output layer includes a global average pooling layer, a fully connected layer and a Softmax layer.

Images