Bearing fault diagnosis method based on multi-scale image Transform

Abstract

The invention relates to the field of bearing fault diagnosis, in particular to a bearing fault diagnosis method based on a multi-scale image Transform, and the method comprises the following steps: S1, obtaining a vibration signal of a bearing; s2, the obtained vibration signals are preprocessed; s3, converting the vibration signal into a time-frequency diagram through wavelet transform; s4, constructing an adjacent matrix of fault feature representation nodes; a step S5 of performing extraction and aggregation on fault features by using a multi-scale image Transform framework; s6, obtaining structural insight through a central coding technology and a spatial coding technology of graph nodes; s7, utilizing a multi-head self-attention mechanism to promote feature extraction and learning of fault nodes; s8, inputting the learned features into a classifier, and carrying out fault classification; a multi-scale feature aggregation module and a graph node centrality coding and space coding technology can be combined for bearing fault diagnosis, so that the fault diagnosis precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of bearing fault diagnosis, and more specifically to a bearing fault diagnosis method based on a multi-scale graph Transformer. Background Art

[0002] Bearings are important components in rotating machinery and are prone to failure in complex operating environments. Traditional fault diagnosis methods rely on signal processing and feature extraction techniques, such as Fourier transform and empirical mode decomposition. These methods work well on small-scale data sets, but have limited feature extraction capabilities on large-scale data sets and cannot fully explore the potential relationships between fault data. Deep learning methods have developed rapidly in recent years, especially graph neural networks (GNNs), which perform well in processing non-Euclidean data. However, existing graph neural network (GNN) models still face problems in fault diagnosis, such as weak single-scale feature extraction capabilities and ignoring the complex relationships between fault data. Summary of the invention

[0003] The purpose of the present invention is to provide a bearing fault diagnosis method based on a multi-scale graph Transformer, which can combine the multi-scale feature aggregation module and the centrality coding of graph nodes with the spatial coding technology for bearing fault diagnosis, thereby improving the fault diagnosis accuracy.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] A bearing fault diagnosis method based on multi-scale graph transformer is proposed. The multi-scale graph transformer framework is used to collect and aggregate the neighborhood information of graph nodes from different scales to enhance the feature representation of the central node. The centrality encoding technology and spatial encoding technology of graph nodes are used to further enrich the node feature information and optimize the structural representation of fault data. The multi-head self-attention mechanism is used for feature extraction, and the feedforward network is combined to achieve efficient fault classification.

[0006] The method comprises the following steps:

[0007] Step S1: obtaining a vibration signal of a bearing;

[0008] Step S2: preprocessing the acquired vibration signal;

[0009] Step S3: converting the vibration signal into a time-frequency diagram through wavelet transform;

[0010] Step S4: construct an adjacency matrix of fault feature representation nodes;

[0011] Step S5: Extract and aggregate fault features using a multi-scale graph Transformer framework;

[0012] Step S6: Obtain structural insights through centrality coding techniques and spatial coding techniques of graph nodes;

[0013] Step S7: Use the multi-head self-attention mechanism to promote feature extraction and learning of fault nodes;

[0014] Step S8: input the learned features into the classifier to perform fault classification;

[0015] The vibration signal preprocessing step includes standardizing the collected data and adopting sliding window sampling, wherein each window contains 1024 sampling points and the sliding step length is 128;

[0016] The wavelet transform converts the vibration signal into a time-frequency diagram with a size of 32×32;

[0017] The adjacency matrix is ​​constructed based on the shortest path distance between nodes, and the k nodes closest to the node itself are selected;

[0018] The multi-scale graph Transformer framework includes a multi-scale feature aggregation module, which is used to broaden the receptive field of the node and enrich the feature representation;

[0019] The multi-scale feature aggregation module enhances the feature representation of the central node by aggregating feature information from neighboring nodes of different scales;

[0020] The centrality encoding technology of the graph nodes embeds the degree of the nodes into the node features to enhance the fault feature representation;

[0021] The spatial encoding technique is used to embed the shortest path distance between nodes to represent the relative positions of graph nodes;

[0022] The multi-head self-attention mechanism improves the ability to extract fault node features by focusing on more important fault features;

[0023] The method was tested on the Case Western Reserve University bearing dataset and rolling bearing experimental device, achieving a peak diagnostic accuracy of 99.86% and consistently maintaining an average accuracy of over 98%;

[0024] The classifier is a Softmax classifier, which is used to classify fault nodes. The fault types include normal state, inner ring fault, outer ring fault and rolling element fault.

[0025] The beneficial effects of the present invention are:

[0026] By combining the multi-scale feature aggregation module, the centrality encoding technology of the graph nodes and the spatial encoding technology, the richness and discrimination ability of the node feature representation are enhanced, which can more accurately identify the fault type of the rolling bearing; by utilizing the multi-head self-attention mechanism, it can adaptively focus on key fault features, effectively resist the interference of noise and complex environmental changes, and improve the stability and robustness of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0028] Figure 1 It is a schematic diagram of the steps of the bearing fault diagnosis method based on the multi-scale graph Transformer of the present invention;

[0029] Figure 2 It is a schematic diagram of the graph convolutional network architecture of the present invention;

[0030] Figure 3 is a schematic diagram of a multi-level feature integration mechanism of the present invention;

[0031] Figure 4 is a schematic diagram of the diagnosis results of different methods in the CWRU data set experiment of the present invention;

[0032] Figure 5 It is a schematic diagram of the confusion matrix of the test results of the CWRU data set experiment of the present invention;

[0033] Figure 6 It is a visualization diagram of the t-distributed stochastic neighbor embedding (t-SNE) results of the present invention. DETAILED DESCRIPTION

[0034] The present invention is further described in detail below with reference to the accompanying drawings.

[0035] like Figures 1 to 6 As shown, in order to achieve the technical effect of "combining the multi-scale feature aggregation module and the centrality encoding of the graph nodes with the spatial encoding technology for bearing fault diagnosis, thereby improving the fault diagnosis accuracy", the steps and functions of a bearing fault diagnosis method based on a multi-scale graph Transformer are described in detail below;

[0036] A bearing fault diagnosis method based on multi-scale graph Transformer, such as Figure 1As shown in the figure, the multi-scale graph transformer Transformer framework is used, which combines a unique multi-scale feature aggregation mechanism of graph nodes, which is designed to broaden the receptive field and enrich the feature representation; structural insights can be obtained through the centrality encoding technology and spatial encoding technology of graph nodes; by utilizing the multi-head self-attention mechanism, effective feature extraction and learning of fault nodes are promoted, and the ability to identify key fault features is significantly enhanced;

[0037] The method comprises the following steps:

[0038] Step S1: obtaining a vibration signal of a bearing;

[0039] Step S2: preprocessing the acquired vibration signal, including standardizing the acquired data and adopting sliding window sampling, where each window contains 1024 sampling points and the sliding step length is 128;

[0040] Step S3: converting the vibration signal into a time-frequency diagram by wavelet transform, wherein the wavelet transform converts the vibration signal into a time-frequency diagram of size 32×32;

[0041] Step S4: construct an adjacency matrix of the fault feature representation node, the adjacency matrix is ​​constructed based on the shortest path distance between nodes, and selects k nodes that are closest to the node itself;

[0042] Step S5: Extract and aggregate fault features using a multi-scale graph Transformer framework. The multi-scale graph Transformer framework includes a multi-scale feature aggregation module. The multi-scale feature aggregation module is used to broaden the receptive field of the node and enrich the feature representation. The multi-scale feature aggregation module enhances the feature representation of the central node by aggregating feature information from neighboring nodes of different scales.

[0043] Step S6: Structural insights are obtained through the centrality encoding technology and spatial encoding technology of graph nodes. The centrality encoding technology of graph nodes embeds the degree of nodes into the node features to enhance the fault feature representation; the spatial encoding technology is used to embed the shortest path distance between nodes to represent the relative positions of graph nodes;

[0044] Step S7: using a multi-head self-attention mechanism to promote feature extraction and learning of fault nodes. The multi-head self-attention mechanism improves the ability to extract fault node features by focusing on more important fault features.

[0045] Step S8: Input the learned features into the classifier for fault classification. The classifier is a Softmax classifier, which is used to classify the fault nodes. The fault types include normal state, inner race fault, outer race fault and rolling element fault. This method is tested on the Case Western Reserve University bearing data set and rolling bearing experimental device, achieving a peak diagnostic accuracy of 99.86% and always maintaining an average accuracy of more than 98%.

[0046] like Figure 2 As shown in the figure, the graph convolution network (GCN) is a deep learning algorithm that can effectively process non-Euclidean data; the graph convolution network (GCN) is a convolutional neural network specifically used for graph operations. The convolution of the graph convolution network (GCN) can be implemented in two different ways: spectrum-based convolution and space-based convolution. The spectral domain graph convolution network (GCN) uses the spectral decomposition of the graph Laplacian operator to capture signal correlation. Specifically, the graph convolution formula is as follows:

[0047] f*x=U((U T f)(U T x))=Ug w U T x

[0048] Where f∈R N and x∈R N denote the convolution kernel and eigenvector respectively. U is the Fourier basis, derived from the eigendecomposition of the Laplace matrix L = UΛU -1 , U and Λ are the eigenvector and eigenvalue matrix L respectively. g w =diag(U T h)=diag(w1,w2,...,w N ) represents a set of learnable parameters. Since the computational cost of directly decomposing the graph is very high, g w Approximate using Chebyshev polynomials:

[0049]

[0050] in, and I n is the identity matrix. By restricting the approximation to first order (K=1) and exploiting the Chebyshev polynomial properties, the graph convolution can be simplified to:

[0051]

[0052] By setting w′=w0=-w1 and further simplifying, the standard GCN forward propagation equation is as follows:

[0053] Z=σ(D -0.5 AD -0.5 (XW))

[0054] in, A is the adjacency matrix of X, is the parameter matrix that W needs to learn, and σ is the activation function;

[0055] Multi-scale feature aggregation (GraphTransformer) module, a novel graph node feature aggregation module is proposed, which effectively utilizes the diverse neighborhood feature information to broaden the receptive field of the node. By aggregating the feature information from neighboring nodes of different scales, it enhances the feature representation of the central node;

[0056] In addition, by combining the centrality coding technology and spatial coding technology of graph nodes and embedding the location information of graph nodes, the richness of node feature information is enhanced;

[0057] like Figure 3 As shown in Figure 1, a schematic diagram of the multi-level feature integration mechanism is depicted. In the graph neural network (GNN) framework, incorporating information from neighboring nodes into node feature representation plays a key role. In order to enhance feature representation, this paper fuses information from different neighborhoods and designs a multi-scale feature aggregation module. For node v, let N k (v) = {u∈v|d(v,u)≤k} represents its k-th order neighborhood, where d(v,u) represents the shortest path distance v between d(v,u) and u. N 0 (v) = {v} is defined as the node itself. In the multi-scale feature aggregation module, N k (v) is defined as the aggregation v of the node neighborhood feature information, expressed as:

[0058]

[0059] Neighborhood feature information It can be obtained by the above formula, where k is fixed as a hyperparameter. Therefore, for each node G in the graph, its corresponding neighborhood feature information can be generated, expressed as X G ∈R n×d×(k+1) Each node in the graph aggregates feature information from its k-scale neighboring nodes. Therefore, the feature matrix G of the node in the graph is expanded to It indicates For the initial feature matrix of the -order neighboring nodes Aggregation k of feature information.

[0060] As mentioned before, the receptive field depends on the value of k. Consider a normalized adjacency matrix and a feature matrix X. In the case of k = 0, the aggregated node features come directly from the nodes themselves, specifically the matrix X. In contrast, when k equals 1, the aggregated features are computed as the average of the directly adjacent nodes, expressed by the matrix multiplication For higher values ​​of k, such as k = 2, the aggregated node features contain both first-order and second-order neighbors, via successive matrix multiplications This trend continues in higher order k. Therefore, the k-th order neighborhood node feature matrix can be expressed as:

[0061]

[0062] The traditional GNN architecture only considers a single-scale information aggregation method, that is, k uses a fixed scale to aggregate information from neighboring nodes. Therefore, this paper introduces a multi-scale feature aggregation module, which aims to integrate different k values, representing the feature information of neighboring nodes of different scales, to enhance the feature representation of nodes. In addition to the inherent feature information of nodes, their structural information also contains valuable content in the graph. This paper adopts two methods, centrality encoding and spatial encoding, to embed the structural information of graph nodes. According to the degree of each node, the centrality information is embedded, which can be directly merged into the features of the node as the input of the Transformer encoder.

[0063]

[0064] For sequential data, there are two ways to embed spatial information: absolute position encoding and relative position encoding. However, for graph data, nodes are not arranged in order but connected by edges. In order to encode the structural information of graph data, this paper proposes the following method: φ(ν i ,ν j ). In the context of a graph G, when two nodes ν i and ν j When directly connected by edges (indicating mutual proximity), the spatial encoding of these nodes is chosen to be the shortest path between them, in which case the shortest path is just the direct edge itself. However, if there is no direct edge connecting ν i and ν j , a default or placeholder value φ is assigned, typically set to -1 to indicate that there is no direct connection. The embedding process accordingly determines the spatial encoding for each pair of nodes in the graph by applying this logic.

[0065]

[0066] Among them, b φ (v i ,v j) is a learnable parameter shared by all layers. After obtaining the feature matrix of the two structural embedding codes, they are input into the multi-scale feature aggregation module to fuse the information of neighboring nodes of different scales. This method effectively solves the problem of insufficient structure mining and feature extraction between nodes of fault vibration signals. First, the collected original bearing fault data is subjected to wavelet transform to obtain the fault time-frequency diagram. Then, by calculating the distance between two fault feature representation nodes, the k fault feature representation nodes that are closest to the node itself are selected to construct an adjacency matrix. The adjacency matrix and the feature representation of the fault node are used as the input of the fault diagnosis model.

[0067] In the fault diagnosis model, the feature matrix and adjacency matrix of the faulty node first pass through the multi-scale feature aggregation module to aggregate the k-scale neighborhood information of the faulty node, denoted as; A linear layer is then used to obtain a low-dimensional embedding representation of the faulty node. A linear layer is then used to obtain a low-dimensional embedding representation of the faulty node.

[0068]

[0069] It is assumed that the dimension of the characteristic matrix of the faulty node is The output dimension of the embedding layer is d out The resulting tensor is then fed into the Transformer’s encoder, which consists of a multi-head self-attention (SA) mechanism and a feed-forward network (FFN). Layer normalization (LN) is applied before each layer. The FFN layer consists of two linear layers and a nonlinear activation function GELU.

[0070]

[0071] After the faulty node is subjected to feature extraction by Transformer, the features are reorganized through the fully connected layer and then accurately classified through the Softmax classifier.

[0072] To verify the effectiveness of the new method in diagnosing rolling element bearing faults, this study combined the use of a benchmark dataset of motor bearings from Case Western Reserve University (CWRU) in the United States. The deep learning framework selected in this study is Pytorch, and the experiments are performed on a Windows 10-based computer system equipped with an AMD R5-3600 processor and a GTX3060 graphics processing unit (GPU) to ensure efficient computational requirements. The bearing fault test bench mainly consists of an induction motor, a torque sensor, and a dynamometer. Deep groove ball bearings SKF6205 and SKF6203 are the objects of the study. Vibration signals are collected under four different operating conditions, namely normal state, inner ring fault, outer ring fault, and rolling element fault. Vibration signals of four fault levels are collected for the rolling element and outer ring, while vibration signals of three fault levels are collected for the outer ring. All vibration signals are collected under motor loads of 0, 1, 2, and 3 horsepower (1 horsepower ≈ 0.735 kilowatts). The sampling frequencies are 12 and 48kHz, and the bearing data are obtained by sampling on the drive end bearing or the fan end bearing. The normal state refers to the absence of any fault.

[0073] In order to extract the time-frequency characteristics of the vibration signal, the original vibration time series is converted into a time-frequency diagram by wavelet transform. When selecting data to generate a time-frequency diagram, the collected data is first standardized. After standardization, a sliding window is used to assist sampling. The sliding window samples the data sequentially with a step size of 128, with 1024 sampling points as the unit. The size of the time-frequency diagram generated by the original vibration time series after wavelet transform is 32×32. The time-frequency diagram has the characteristics of both time domain and frequency domain, which respectively amplifies the relevant details of time and frequency, so that the model can more effectively extract the fault signal characteristics. In order to simulate the real scene, this paper divides the collected data into 10 categories according to the fault type, and each type of fault contains 100 time-frequency diagram samples. 60% of the data is used as a training set, 20% as a test set, and 20% as a validation set. The experiment sets the number of Transformer layers to 5, the embedding layer dimension to 512, the initial learning rate to 0.0001, the optimizer to Adam, the iteration batch count threshold to 100, and the Dropout to 0.5 to prevent overfitting.

[0074] To verify the effectiveness and advantages of our proposed method, we built six different fault diagnosis models for comparison. These models include a standard graph convolutional network (GCN), and a Chebyshev graph convolutional network (ChebyNet) with 1, 2, and 3 scale receptive fields. In addition, we also compared our method with a traditional 5-layer convolutional neural network (CNN) and the model Graphormer. In order to mitigate the impact of randomness on our research results, we conducted five independent experiments to ensure the robustness of our proposed method. The main evaluation criterion is the overall classification accuracy, including the highest, lowest, and average accuracy in five iterations.

[0075] like Figure 4 As shown in the figure, the diagnostic results of different methods in the CWRU dataset experiment, it is obvious that MSGraphormer outperforms the other seven methods in terms of stability. In five experiments, MSGraphormer achieved accuracies of 99.66%, 99.55%, 99.49%, 99.67% and 99.86%, respectively. In addition, its average accuracy is 99.56% and the minimum standard deviation is 0.12%, highlighting the excellent stability of the proposed method. The order and accuracy of the Chebyshev polynomial of the Chebyshev graph convolutional network (ChebyNet) are positively correlated, indicating that in the graph convolutional network (GCN), increasing the receptive field scale can improve the accuracy. MSGraphormer combines receptive fields of multiple scales, which can not only aggregate information from neighboring nodes of different receptive fields, but also integrate this information to produce more discriminative features, so it has a better classification effect.

[0076] like Figure 5 As shown in the figure, the confusion matrix diagram can show the classification results more intuitively.

[0077] like Figure 6 As shown in the figure, in order to further highlight the superiority of the method proposed in this paper, the t-distributed random neighbor embedding (t-SNE) visualization results are given. It can be seen from the figure that the method proposed in this paper has shown excellent ability in classifying ten different fault types. This ability mainly comes from the ability of this method to extract fault feature information from multiple scales, thereby enhancing the description ability of features and improving the utilization of relevant information. In addition, the addition of Transformer technology enables the model to adaptively learn the representation ability of fault node features, thereby further improving the overall performance of the classification system.

[0078] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A bearing fault diagnosis method based on multi-scale graph Transformer, characterized by: The method comprises the following steps: Step S1: obtaining a vibration signal of a bearing; Step S2: preprocessing the acquired vibration signal; Step S3: converting the vibration signal into a time-frequency diagram through wavelet transform; Step S4: construct an adjacency matrix of fault feature representation nodes; Step S5: Extract and aggregate fault features using a multi-scale graph Transformer framework; Step S6: Obtain structural insights through centrality coding techniques and spatial coding techniques of graph nodes; Step S7: Use the multi-head self-attention mechanism to promote feature extraction and learning of fault nodes; Step S8: Input the learned features into the classifier for fault classification.

2. The bearing fault diagnosis method based on multi-scale graph Transformer according to claim 1 is characterized in that: The vibration signal preprocessing step includes standardizing the collected data and adopting sliding window sampling, wherein each window contains 1024 sampling points and the sliding step length is 128.

3. The bearing fault diagnosis method based on multi-scale graph Transformer according to claim 1 is characterized in that: The wavelet transform converts the vibration signal into a time-frequency diagram with a size of 32×32.

4. The bearing fault diagnosis method based on multi-scale graph Transformer according to claim 1 is characterized in that: The adjacency matrix is ​​constructed based on the shortest path distance between nodes, and the k nodes closest to the node itself are selected.

5. The bearing fault diagnosis method based on multi-scale graph Transformer according to claim 1 is characterized in that: The multi-scale graph Transformer framework includes a multi-scale feature aggregation module, which is used to broaden the receptive field of the node and enrich the feature representation.

6. The bearing fault diagnosis method based on multi-scale graph Transformer according to claim 5 is characterized by: The multi-scale feature aggregation module enhances the feature representation of the central node by aggregating feature information from neighboring nodes of different scales.

7. The bearing fault diagnosis method based on multi-scale graph Transformer according to claim 1 is characterized in that: The centrality encoding technology of the graph nodes embeds the degree of the nodes into the node features to enhance the fault feature representation.

8. The bearing fault diagnosis method based on multi-scale graph Transformer according to claim 1 is characterized in that: The spatial encoding technique is used to embed the shortest path distance between nodes to represent the relative positions of graph nodes.

9. The bearing fault diagnosis method based on multi-scale graph Transformer according to claim 1 is characterized in that: The multi-head self-attention mechanism improves the ability to extract fault node features by focusing on fault features.

10. The bearing fault diagnosis method based on multi-scale graph Transformer according to claim 1, characterized in that: The classifier is a Softmax classifier, which is used to classify fault nodes. The fault types include normal state, inner ring fault, outer ring fault and rolling element fault.

Images