Bearing fault diagnosis method and system

Abstract

The present application provides a bearing fault diagnosis method and system, which provides an MCHformer architecture that integrates CNN and hybrid Transformer that fuse global and local information, and can combine the advantages of CNN and transformer: first, local features of different scales are obtained through dilated convolution extraction, and then long-range dependencies are established through the feature extractor of the hybrid transformer structure. While maintaining local features, strong and robust global features in the context are extracted, and spatial features of different scales are fused. Finally, the fused features are input into the newly designed structure for adaptive learning to obtain fault category diagnosis results with high accuracy in a high noise environment. The transformer provided by the present application can effectively reduce the amount of information used and better focus on adjacent points. The neural network model of the present application can directly process noise and vibration data, and can effectively diagnose the type of fault without the use of advanced preprocessing technology.

Description

Technical Field

[0001] The present application relates to the technical field of mechanical fault diagnosis, and in particular to a bearing fault diagnosis method and system. Background Art

[0002] With the rapid development of society, bearings have been widely used in rotating machinery and equipment such as rail transportation, aerospace vehicles, automobiles, and industrial robots. The operating environment of these equipment is relatively complex, such as high temperature and high pressure, variable speed, and high humidity. Bearings are prone to failure, which will cause huge economic losses. Therefore, it is very important to carry out monitoring and diagnosis of bearings, reduce the occurrence rate of failures, and ensure the safe and reliable operation of machinery.

[0003] Data-driven bearing fault diagnosis methods do not require complex model building. Instead, they leverage large amounts of data to build and continuously optimize models, resulting in excellent diagnostic results. This makes them ideal for intelligent fault diagnosis of complex machinery. Currently, data-driven fault diagnosis methods can be generally categorized into intelligent diagnosis methods based on traditional machine learning, intelligent diagnosis methods based on deep learning, and fault diagnosis methods based on deep transfer learning.

[0004] With the rapid development of deep learning, it has been widely used due to its powerful feature extraction capabilities. Researchers use convolutional neural networks and their improved models for fault diagnosis. Shao et al. proposed a new method called ensemble deep autoencoder (EDAES) for intelligent fault diagnosis of rolling bearings. Li et al. proposed a new fault diagnosis method based on adaptive multiscale morphological filtering (AMMF) and modified hierarchical permutation entropy (MHPE) to identify different health conditions of planetary gearboxes. Chen et al. explored the random forest (RF) ensemble learning algorithm for the detection and diagnosis of early faults (including line faults, degradation, open circuits, and partial shading) of photovoltaic arrays. Combining multiple learning algorithms, they achieved superior diagnostic performance. Zhao et al. proposed an analog circuit fault diagnosis method based on deep belief networks (DBNs). Chen et al. attempted to summarize the recent research and development trends in the field of basic research on mechanical fault diagnosis from four main aspects: fault mechanism, sensor technology and signal acquisition, signal processing, and intelligent diagnosis. Zhu et al. proposed a fault diagnosis method consisting of three parts: (1) effectively analyzing the output characteristics and electrical characteristic vector distribution of photovoltaic arrays under typical fault conditions; (2) introducing the unit method and Gaussian kernel function into the fuzzy mean algorithm to improve the applicability of unsupervised screening for various fault samples and the fuzzy clustering ability; (3) using clustered data as input to establish a probabilistic neural network fault diagnosis model. Li et al. proposed a new rolling bearing fault diagnosis method based on deep learning. Traditional composite fault diagnosis methods focus on the fault diagnosis of bearings or gears. In the fault diagnosis process, composite faults are always regarded as an independent fault mode, and the relationship between single faults and composite faults is not fully considered. To solve this problem, Huang et al. proposed an intelligent composite fault diagnosis method called deep decoupled convolutional neural network. Xu et al. gave the advantages of fuzzy neural network technology in equipment fault diagnosis and explained the basis of fuzzy theory and neural network technology. Wenkui et al. proposed a new fault diagnosis method based on convolutional neural network (CNN).

[0005] However, these studies failed to account for the strong noise interference present in industrial production. Due to the high noise levels in real-world environments, the vibration signals collected by sensors are non-stationary, resulting in low-quality features extracted by traditional neural networks. Furthermore, due to the strong noise interference, the periodic characteristics of the bearing vibration signals can vary significantly, making it difficult for CNNs to extract high-quality features and unable to resolve long-range dependencies. Summary of the Invention

[0006] This application addresses the shortcomings of existing technologies and provides a bearing fault diagnosis method and system. This application addresses the problem that most models cannot fully capture the multi-scale features in the original vibration signal and the problem of feature extraction and fault diagnosis of machinery in a noisy environment. This application proposes a neural network based on the fusion of a multi-scale CNN and a hybrid transformer, which is based on global and local joints. This method can effectively improve the fault identification capability of the convolutional neural network and maintain a high accuracy rate under strong interference noise. This application specifically adopts the following technical solutions.

[0007] First, to achieve the above objectives, a bearing fault diagnosis method is proposed, which includes the following steps: first, collecting bearing vibration data, performing dilated convolution on the vibration data, and obtaining features of different scales; second, extracting global information from features of different scales through a feature extractor of a hybrid Transformer architecture, establishing long-range dependencies, and fusing spatial features of different scales; third, inputting the fused features into a neural network model that has been trained through adaptive learning to diagnose and identify the bearing fault category.

[0008] Optionally, a bearing fault diagnosis method as described in any of the above, wherein the feature extractor of the hybrid Transformer architecture includes two different transformers, one Transformer adopts a multi-scale convolutional attention mechanism (MSCA) and an inverse residual feedforward network (IRFFN), and the other Transformer adopts an SE attention mechanism and an inverse residual feedforward network (IRFFN), and the two transformers are connected in series. Compared with the traditional Transformer with only the attention mechanism as the main component, the bearing fault diagnosis method introduces a fusion method of traditional convolutional neural networks and Transformers, and ensures the richness of features through the introduction of multi-scale convolution and multiple attention mechanisms, thereby improving the discriminability of the diagnosis method for features.

[0009] Optionally, a bearing fault diagnosis method as described in any of the above, wherein a multi-scale convolutional attention mechanism (MSCA) is used in the Transformer to replace the multi-head attention mechanism in the original Transformer, and by introducing multiple scales, it uses a depth-separable convolution module to replace the traditional convolution block to better extract local information, and the addition of a residual structure can better be used to extract global context information, and finally correlation modeling is performed in the channel dimension through convolution with a kernel size of 1x1.

[0010] Optionally, in a bearing fault diagnosis method as described above, the convolution kernels of each DW convolution module of the multi-scale convolution attention mechanism (MSCA) are: 5x5, 1x7, 1x11, 1x21, 7x1, 11x1, 21x1 and 1x1; each DW convolution module uses convolution kernels of different sizes to extract multi-scale features, and uses them as attention weights to reweight the input of MSCA.

[0011] Optionally, in any of the above bearing fault diagnosis methods, two different Transformers replace the multi-head attention mechanism in the original Transformer with a multi-scale convolutional attention mechanism (MSCA) and an SE attention mechanism (Attention), which can better extract feature information by separating the input matrix X and the matrix W. K 、W V 、W Q Multiply them to get K, Q, and V, then perform dot product on Q and K to get the softmax layer. Calculate the similarity or correlation between Q and K to get the weight coefficient. Then perform dot product on the softmax layer and V, and use the weight coefficient of the softmax layer to perform weighted summation on V to get the output of the self-attention module.

[0012] Optionally, in a bearing fault diagnosis method as described in any of the above, the inverse residual feedforward network (IRFFN) is used to replace the original MLP layer of the transformer, which consists of an expansion layer, a depth convolution and a projection layer, and is used to change the position of the residual connection; the inverse residual feedforward network (IRFFN) uses a DW convolution layer to replace the traditional convolution, introduces a GELU activation function to replace the ELU activation function, and adds a residual structure to eliminate the effects caused by gradient vanishing and gradient exploding.

[0013] Optionally, a bearing fault diagnosis method as described above, wherein the inverse residual feedforward network (IRFFN) has the following expression: IRFFN(X)=Conv(F(Conv(X))), F(X)=DWConv(X)+X, wherein the convolution kernel of the depthwise separable convolution DWConv(X) is 3x3, which is used to extract local information.

[0014] Optionally, in a bearing fault diagnosis method as described in any of the above, the SE attention mechanism is used to replace the multi-head attention mechanism of the original Transformer, first globally pooling the input features to convert the features into W×H×C, and then adjusting the W×H×C features through a fully connected layer and an activation function to convert them into the weight of each feature, which is then multiplied with the input features.

[0015] A bearing fault diagnosis system based on any of the methods described above, comprising: a sensing module for collecting vibration data of the bearing; a feature extraction module for performing dilation convolution on the vibration data to obtain features of different scales, and then extracting global information from the features of different scales through a feature extractor of a hybrid Transformer architecture, establishing long-range dependencies, and fusing spatial features of different scales; a fault identification module for receiving the fused features and diagnosing and identifying the fault category of the bearing through a neural network model trained through adaptive learning.

[0016] Optionally, in a bearing fault diagnosis system as described above, the feature extraction module adopts a fault diagnosis architecture that integrates CNN and hybrid Transformer that fuse global and local information, performs feature extraction on the grayscale image obtained by processing the vibration data through MCHformer-RDP, obtains fault features, and then classifies the fault according to the feature information.

[0017] Beneficial effects

[0018] The bearing fault diagnosis method and system provided by the present application provides an MCHformer architecture that integrates CNN and hybrid Transformer that fuse global and local information, which can combine the advantages of CNN and transformer: first, local features of different scales are obtained through dilated convolution extraction, and then long-range dependencies are established through the feature extractor of the hybrid transformer structure. While maintaining local features, strong and robust global features in the context are extracted, and spatial features of different scales are fused. Finally, the fused features are input into the newly designed structure for adaptive learning to obtain fault category diagnosis results with high accuracy in a high noise environment. The transformer provided by the present application can effectively reduce the amount of information used and better focus on adjacent points. The neural network model of the present application can directly process noise and vibration data, and can effectively diagnose the type of fault without the use of advanced preprocessing technology.

[0019] Specifically, this application designs the MCHformer framework by combining the advantages of CNN and Transformer. CNN uses the property of translation invariance to obtain local features, while Transformer extracts global effective features by establishing long-range dependencies. By fusing the spatial features of two different scales, it can collect more representative features that are better suited for strong noise conditions.

[0020] In order to extract high diagnostic quality and identify multi-scale features, this application designs two different Transformer framework networks and two attention mechanisms, which can effectively reduce the extraction of redundant information and thus enhance the diagnostic effectiveness.

[0021] This application creatively combines MCHformer and R-drop to develop MCHformer-RDP for noise interference fault diagnosis, so that all parameters of the model are subject to equal regularization constraints as much as possible, which can effectively prevent overfitting.

[0022] The neural network model proposed in this application can directly process noise and vibration data without the need for advanced preprocessing technology. Through analysis of experimental data, the technical solution of this application is effective and feasible, and the diagnostic and recognition effect is significantly better than the existing technology.

[0023] Other features and advantages of the present application will be set forth in the following description, and in part will be apparent from the description, or may be learned by practicing the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings are used to provide a further understanding of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings:

[0025] Figure 1 It is a schematic diagram of the principle of the existing convolutional neural network CNN;

[0026] Figure 2 This is a schematic diagram of the principle of the multi-head self-attention mechanism module (MHSA) in the existing transformer;

[0027] Figure 3 It is a schematic diagram of the principle model of the existing R-Drop technology;

[0028] Figure 4 This is the principle block diagram of the 4CHformer model provided in this application;

[0029] Figure 5 This is a block diagram of the principle of the multi-scale convolutional attention mechanism (MSCA) used in this application;

[0030] Figure 6 This is a block diagram of the principle of the attention mechanism used in this application;

[0031] Figure 7 This is a block diagram of the principle of the Inverse Residual Feedforward Network (IRFFN) used in this application;

[0032] Figure 8 This is a block diagram of the SE attention mechanism (SE) used in this application;

[0033] Figure 9This is a diagram of the equipment at the Case Western Reserve University (CWRU) test bed used in this application;

[0034] Figure 10 This is a comparison chart of the denoising capabilities of the MCHformer architecture used in this application in different noise environments;

[0035] Figure 11 The t-SNE graph obtained by visualizing the diagnosis results of the MCHformer model under SNR 5 in case A using the t-SNE algorithm, where (a) Proposed method; (b) VGG; (c) MobileNetV3; (d) ShuffleV2;

[0036] Figure 12 Confusion matrix of the diagnosis results of the MCHformer model, where (a) Proposed method; (b) VGG; (c) MobileNetV3; (d) ShuffleV2;

[0037] Figure 13 Schematic diagram of the XJTU experimental platform used in this application;

[0038] Figure 14 A comparison graph of the denoising capabilities of different models;

[0039] Figure 15 t-SNE graph of the recommended model, where (a) Proposed method; (b) VGG; (c) MobileNetV3; (d) ShuffleV2

[0040] Figure 16 Confusion matrix of the recommended model, where (a) Proposed method; (b) VGG; (c) MobileNetV3; (d) ShuffleV2. DETAILED DESCRIPTION

[0041] To make the purpose and technical solutions of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be clearly and completely described below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0042] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as such herein, will not be interpreted in an idealized or overly formal sense.

[0043] The term “connection” as used in this application may mean a direct connection between components or an indirect connection between components via other components.

[0044] Traditional convolutional neural networks (CNNs) such as Figure 1 As shown in , it is famous for its efficient feature extraction capabilities. It usually consists of convolutional layers, pooling layers, activation layers, and fully connected layers, such as Figure 1 As shown in the figure, the parameters of the same convolution kernel are shared through translation invariance, local features of different scales are extracted, and back propagation is used to update the parameters to achieve continuous optimization of the model.

[0045] Considering that a key factor affecting CNN's ability to learn long-range dependencies is the length of the path that forward and backward signals in the network must pass through, and the Transformer structure can well solve the long-range dependency problem, in recent years, some people have gradually tried to apply the Transformer architecture to the field of fault diagnosis.

[0046] The traditional Transformer architecture can be referred to Figure 2 As shown in , the transformer is a multi-layer network based on the superposition of transformer blocks, where the transformer block is mainly composed of a multi-head self-attention mechanism module (MHSA), a position feedforward module (FNN), a layer normalization module (LayerNorm), a residual connection, etc. Figure 2 As shown in Figure 1, the core of the Transformer is Multi-HeadAttention, which consists of multiple Self-Attentions. The matrices Q, K, and V are required for calculation. Q, K, and V are obtained by linearly transforming the input of the Self-Attention. The formula is shown in (1).

[0047]

[0048] In existing research, Ding et al. proposed a novel time-frequency transformer (TFT) model inspired by the remarkable success of the vanilla transformer in sequence processing. Wu et al. proposed an advanced deep learning-based approach that employs a transformer architecture to diagnose simultaneous faults using time series data. This approach can be directly applied to transient data while maintaining accuracy, without the need for a steady-state detector, enabling fault diagnosis in its early stages. Alexakos et al. proposed a combined short-time Fourier transform and image classification transformer model for motor rolling element bearing fault diagnosis. The success of transformers is due to the use of large-scale data, but existing transformer technology is often limited in fault sample size for fault diagnosis tasks. Since the number of bearing fault samples is quite small, it is difficult to support effective training of existing Transformer architectures. Furthermore, existing Transformer architectures lack the translation invariance and locality properties of CNNs, and therefore cannot generalize well to new tasks when data is insufficient.

[0049] Based on the above dilemma, this application combines the unique advantages of CNN and transformer in feature extraction, and proposes a bearing fault diagnosis framework MCHformer-RDP: a global and local joint CNN and hybrid transformer fusion. It can extract useful features from complex vibration signals under strong noise conditions and ignore useless features. This application increases the robustness of the network by collecting local area features and global features. The design concept of MCHformer is to obtain a variety of spatial features including long-term dependencies and local dependencies, and fuse them. In order to ensure the richness and globality of feature scales, two different transformers are designed, including MSCA and Attention, and IRFFN and SE are used for two different transformers respectively, and the two transformer structures are connected in series. The system first obtains features of three different scales through a multi-branch structure, and then extracts global information of these three different scale features through a hybrid transformer and establishes long-range dependencies. In addition, this application also modifies some properties of the transformer to reduce redundant information and better focus on adjacent points.

[0050] Specific reference Figure 4 In the 4CHformer used in this application, IRFFN: The IRFFN module composed of an inverted residual feedforward network replaces the original MLP layer of the transformer. The inverted residual feedforward network (IRFFN) is similar to the inverted residual block and consists of an expansion layer, a depthwise convolution, and a projection layer. Specifically, by changing the position of the residual connection to achieve better performance, the expression is as follows:

[0051] IRFFN(X)=Conv(F(Conv(X))) (4)

[0052] F(X)=DWConv(X)+X (5)

[0053] The activation layer and batch normalization linear layer are omitted. Depthwise separable convolution is used to extract local information, with negligible computational cost. The residual network can improve the ability of gradients to propagate across layers. Experiments have shown that this shortcut helps the network achieve better results.

[0054] MSCA: Figure 5 As shown, convolutional attention is a more effective and efficient way to encode contextual information than the self-attention mechanism in the Transformer. The MSCA module adopts the traditional convolutional module design but introduces multi-scale convolutional attention to further extract global contextual information. DW convolution is used to aggregate local information, multi-branch DW convolution is used to capture multi-scale contextual information, and 1x1 convolution is used to model correlation in the channel dimension.

[0055] SA self-attention mechanism: Self-attention mechanism such as Figure 6 As shown, the input matrix X and the matrix W are K 、W V 、W Q Multiply to get K, Q, and V, then dot-multiply Q and K, and then dot-multiply with V in the softmax layer. Calculate the similarity or correlation between Q and K to get the weight coefficient, and then perform a weighted summation on V to get the output of the self-attention module. Mathematically expressed as (6):

[0056] Attention(Q,K,V)=V·softmax(K T Q) (6)

[0057] Therefore, in order to improve the model's ability to mine features under strong noise conditions and improve the accuracy of fault diagnosis, this application provides a bearing fault diagnosis method and a system using this method through an MCHformer architecture that integrates CNN and hybrid Transformer that fuses global and local information, combining the advantages of CNN and transformer.

[0058] The system collects bearing vibration data through the sensing module; performs dilation and convolution on the vibration data through the feature extraction module to obtain features of different scales; then uses the feature extractor of the hybrid Transformer architecture to extract global information from features of different scales, establish long-range dependencies, and fuse spatial features of different scales; then receives the fused features through the fault identification module, and diagnoses and identifies the bearing fault category through the adaptive learning and trained neural network model.

[0059] Specifically, the hybrid transformer structure used in the system of this application includes MSCA and Attention as well as IRFFN and SE. First, features of three different scales are obtained through a multi-branch structure, and then the hybrid transformer is used to extract global information of the three different scale features and establish long-range dependencies. In addition, this application also modifies some properties of the transformer to reduce redundant information and better focus on adjacent points. The neural network model proposed in this application can directly process noisy vibration data without the use of advanced preprocessing technology. The effectiveness of the method of this application can be effectively verified through the following analysis of the experimental data.

[0060] The robustness of the MCHformer model provided in this application is verified by two bearing fault data sets, and the diagnostic accuracy of the model in different noise environments is discussed.

[0061] In the bearing fault data set, Poisson noise, salt and pepper noise, and Gaussian white noise with different signal-to-noise ratios are added to the original signal to obtain composite signals with different signal-to-noise ratios. The signal-to-noise ratio is expressed as (7), where P s is the power of the signal, P n is the power of the noise, and lg is the logarithm to the base 10.

[0062]

[0063] To verify the effectiveness of the proposed method, fault diagnosis was performed using the CWRU and XJTU datasets. The experimental hardware consisted of an R5-2600X CPU, 16GB of RAM, and an RTX2080ti GPU. The datasets were generated using MATLAB, and the neural network architecture was developed using PyTorch, using Python 3.6.

[0064] First verification test

[0065] This experiment uses the CWRU dataset for fault diagnosis. The dataset is divided into four categories based on the speed of 1730rpm, 1750rpm, 1772rpm, and 1797rpm. Figure 9The following is a diagram of the CWRU test bench equipment. During the experiment, vibration data was collected using an accelerometer at a sampling frequency of 12 kHz. In this experiment, the CWRU dataset was divided into nine fault categories: inner race, outer race, and rolling element faults, with fault diameters of 0.007 mm, 0.014 mm, and 0.021 mm, respectively. Each category contained 1000 samples. The training and test sets for each category were split in a 9:1 ratio. The dataset classification is shown in Table 1.

[0066] Table 1 Composition of the CWRU dataset

[0067]

[0068] In order to verify the superiority of MCHformer in different noise environments, this application verifies the denoising ability of the model on a composite signal dataset with signal-to-noise ratios of -5, 0, and 5, and obtains the results in Table 2 and Figure 10 The data is used to compare their denoising capabilities:

[0069] Table 2 Comparison of denoising capabilities of different models on the CWRU dataset

[0070] Table 2 (a) SNR = -5

[0071]

[0072] Table 2(b) SNR=0

[0073]

[0074] Table 2(c) SNR=5

[0075]

[0076] From Table 2 and Figure 10 As can be seen, MCHformer demonstrates strong feature mining capabilities across diverse complex noise conditions. In a noisy environment with a signal-to-noise ratio (SNR) of 0dB, the diagnostic accuracy of the comparison model is lower. However, the average accuracy of MCHformer across the four datasets is 28.92% higher than that of VGG, 26.67% higher than that of the Shuffle_v2 model, and 22.56% higher than that of the MobileNetV3 model. The accuracy of the comparison model on datasets A, B, C, and D, under three different SNR conditions, is over 20% lower than that of the recommended model. Figure 11The t-SNE algorithm was used to visualize the diagnosis results of the MCHformer model in case A with an SNR of 5. The results show that MCHformer can extract effective features and achieve high-precision fault diagnosis in a strong noise environment. To better evaluate the classification accuracy of MCHformer for different fault types, the diagnosis results are visualized in the form of a confusion matrix. Figure 12 The confusion matrix of the MCHformer model's diagnostic results is shown. The numbers in the dark blue background represent correctly predicted samples, while the numbers in the light blue squares represent incorrectly predicted samples. Ablation experiments based on different signal-to-noise ratios demonstrate the importance and effectiveness of the multi-scale CNN and hybrid transformer architectures. This is attributed to the extraction of neighborhood information at different scales and adaptive learning, which enables MCHformer to mine effective features from signals even in noisy environments.

[0077] Second verification test

[0078] In this experiment, the XJTU-SY bearing fault dataset is used to verify the generalization ability of the MCHformer model.

[0079] Figure 13 The figure shows the XJTU bearing test rig. The bearing test rig consists of an AC induction motor, a motor speed controller, a support shaft, two support bearings (heavy-loaded roller bearings), and a hydraulic loading system. This dataset contains operational fault data for nine categories of rolling bearings. The dataset is categorized by speed and fault type, as shown in Table 3. The training and test sets are automatically split at a ratio of 9:1.

[0080] Table 3 Composition of the XJTU dataset

[0081]

[0082] In this experiment, we try to verify whether MCHformer can mine effective features in signals in a strong noise environment and whether it has good generalization ability. The network is tested on a composite signal dataset with a signal-to-noise ratio of -2 to 12dB. Figure 14The feature extraction capabilities of these models were compared under strong noise conditions tested on the XJ dataset. As shown in the table, MCHformer exhibits high feature extraction capabilities across various noise environments. When the signal-to-noise ratio (SNR) is 12, MCHformer achieves the highest diagnostic accuracy, exceeding 90%. When the SNR is at other levels, MCHformer's accuracy does not exceed 90%, and none of the comparison models achieve an accuracy above 90%. The average accuracy of the MCHformer model across various noise conditions is 87.04%, which is 12.61% higher than Vgg, 16.02% higher than the lowest model, Shuffle_v2, and 13.88% higher than the classic MobileNetV3 model. To better evaluate the MCHformer model's feature extraction capabilities under strong noise conditions on the XJ dataset, the diagnostic results are visualized using t-SNE plots and confusion matrices. Figure 15 and Figure 16 The t-SNE plots and confusion matrices from different models are shown with a signal-to-noise ratio of 6. Table 4 gives the numerical ID numbers for the bearing states. Figure 15 As shown in , it can be observed that there is a clear distance between the distribution of various types of faults, the distribution within a class is relatively dense, and the class recognition is high, which verifies the model feature extraction capability. Figure 16 As shown, MCHformer can mine effective features in strong noise environments, achieving high accuracy in fault diagnosis. When the bearing state is 35Hz / 12kN and a cage fault occurs, Shuffle_v2 has the lowest recall rate of 93.33%. This is because the characteristic signal of the cage fault is relatively weak and easily overwhelmed by the low signal-to-noise ratio noise signal. The MCHformer model has the highest recall rate of 100%.

[0083] Table 4 Comparison of denoising capabilities of different models on the XJ dataset

[0084]

[0085] The above is only an embodiment of the present application, and its description is relatively specific and detailed, but it should not be understood as limiting the scope of the patent application. It should be pointed out that for those skilled in the art, various modifications and improvements can be made without departing from the concept of the present application, and these modifications and improvements are all within the scope of protection of the present application.

Claims

1. A bearing fault diagnosis method, characterized in that the steps include: The first step is to collect bearing vibration data and perform dilation convolution on the vibration data to obtain features at different scales. In the second step, the feature extractor of the hybrid Transformer architecture extracts global information from features at different scales, establishes long-range dependencies, and fuses spatial features at different scales. The third step is to input the fused features into the neural network model trained by adaptive learning to diagnose and identify the bearing fault category. The feature extractor of the hybrid Transformer architecture includes two different transformers, one of which adopts a multi-scale convolutional attention mechanism (MSCA) and an inverse residual feedforward network (IRFFN), and the other Transformer adopts a SE attention mechanism and an inverse residual feedforward network (IRFFN). The two transformers are connected in series.

2. The bearing fault diagnosis method according to claim 1, characterized in that: The described Transformer adopts a multi-scale convolutional attention mechanism (MSCA) to replace the multi-head attention mechanism in the original Transformer. It adopts a depth-wise separable convolution module to replace the traditional convolution block to extract local information, and adds a residual structure to extract global context information. Finally, correlation modeling is performed in the channel dimension through convolution with a kernel size of 1x1.

3. The bearing fault diagnosis method according to claim 2, characterized in that: The convolution kernels of each DW convolution module of the multi-scale convolution attention mechanism (MSCA) are: 5x5, 1x7, 1x11, 1x21, 7x1, 11x1, 21x1 and 1x1; Each DW convolution module uses convolution kernels of different sizes to extract multi-scale features and uses them as attention weights to reweight the input of MSCA.

4. The bearing fault diagnosis method according to claim 1, wherein: The two different Transformers replace the multi-head attention mechanism in the original Transformer with the multi-scale convolutional attention mechanism (MSCA) and the SE attention mechanism (Attention), respectively, by separating the input matrix X and the matrix W K 、W V 、W Q Multiply them to get K, Q, and V, then perform dot product on Q and K to get the softmax layer. Calculate the similarity or correlation between Q and K to get the weight coefficient. Then perform dot product on the softmax layer and V, and use the weight coefficient of the softmax layer to perform weighted summation on V to get the output of the self-attention module.

5. The bearing fault diagnosis method according to claim 1, wherein: The Inverse Residual Feedforward Network (IRFFN) is used to replace the original MLP layer of the transformer. It consists of an expansion layer, a depthwise convolution, and a projection layer to change the position of the residual connection. The Inverse Residual Feedforward Network (IRFFN) uses a DW convolution layer to replace the traditional convolution, introduces the GELU activation function to replace the ELU activation function, and adds a residual structure to eliminate the effects of gradient vanishing and gradient exploding.

6. The bearing fault diagnosis method according to claim 5, characterized in that: The Inverse Residual Feedforward Network (IRFFN) has the following expression: IRFFN(X)=Conv(F(Conv(X))), F(X)=DWConv(X)+X, Among them, the convolution kernel of the depth-wise separable convolution DWConv(X) is 3x3, which is used to extract local information.

7. The bearing fault diagnosis method according to claim 1, characterized in that: The SE attention mechanism is used to replace the original Transformer's multi-head attention mechanism. It first performs global pooling on the input features to convert the features into W×H×C. Then, through the fully connected layer and activation function, the W×H×C features are adjusted to become the weight of each feature, which is then multiplied with the input features.

8. A bearing fault diagnosis system based on the method according to any one of claims 1 to 7, characterized in that: include: A sensor module is used to collect vibration data of the bearing; The feature extraction module is used to perform dilated convolution on the vibration data to obtain features at different scales. It then uses a feature extractor based on a hybrid Transformer architecture to extract global information from features at different scales, establish long-range dependencies, and fuse spatial features at different scales. The fault identification module is used to receive the fused features and diagnose and identify the bearing fault category through the adaptive learning and trained neural network model.

9. The bearing fault diagnosis system according to claim 8, characterized in that: The feature extraction module adopts a fault diagnosis architecture that integrates CNN and hybrid Transformer that fuse global and local information. It extracts features from the grayscale image obtained by processing vibration data through MCHformer-RDP to obtain fault features, and then classifies the faults based on the feature information.

Images