Fault diagnosis method for multi-mode hyperperceptual graph Transform
Through the multimodal hyper-perceptual graph Transformer model, the problem of insufficient information redundancy and fine-grained capture in the bearing fault diagnosis is solved, and higher diagnostic accuracy and robustness are achieved.
Patent Information
- Application Number
- CN202510490315.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-18
- Publication Date
- 2025-08-01
AI Technical Summary
The existing graph convolutional neural networks are difficult to effectively capture fine-grained fault modes and higher-order nonlinear interactions in bearing fault diagnosis, and the receiving field size is limited, resulting in information redundancy and insufficient diagnostic accuracy.
The multimodal hyper-perception graph Transformer model is adopted to process bearing data through wavelet transform noise reduction, build an adjacency matrix and divide the data set using the k-nearest neighbor method, combine the hyper-perception expansion residual module, element-level Transformer and attribute convolution to enhance feature extraction capabilities and capture the spatial dependence relationship of multimodal data.
Improve the accuracy and robustness of bearing fault diagnosis, especially in the conditions of noise and vibration waveform changes, maintaining good diagnostic performance.
Smart Images

Figure CN120408349A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of bearing fault diagnosis and graph neural network technology, specifically a fault diagnosis method based on a multi-modal ultra-perceptive graph Transformer. Background Art
[0002] As one of the key components in various rotating machinery, rolling bearings play an extremely important role in mechanical systems such as power turbine drive systems. However, the failure probability of rolling bearings is much higher than that of other components because bearings are often exposed to heavy loads and challenging operating conditions. Therefore, the long-term reliability and real-time health monitoring of bearings are of great significance. The rapid development of deep learning technology has provided new ideas for fault diagnosis. However, standard deep learning models have limited ability to represent complex relational data, while graph convolutional neural networks based on graph-structured data have shown significant advantages in processing complex relationships and non-Euclidean structures. In fault diagnosis, graph convolutional neural networks have certain limitations in terms of the receptive field size. An overly large receptive field may introduce too much irrelevant information, resulting in redundancy. Therefore, the size of the receptive field is usually restricted, which limits the ability of graph convolutional neural networks to fully perceive and extract global information. In addition, graph convolutional neural networks also have certain defects in capturing the spatial dependence relationships and interaction mechanisms between fault data. It is difficult for them to accurately capture fine-grained fault patterns and to model the high-order non-linear interactions between data points. To solve these problems, the present invention proposes a fault diagnosis method based on a multi-modal ultra-perceptive graph Transformer. Summary of the Invention
[0003] In order to better diagnose mechanical faults, the present invention constructs a multi-modal ultra-perceptive graph Transformer model based on the Paderborn bearing data and verifies its effectiveness in experiments, thereby having higher diagnostic accuracy. The technical solution adopted by the present invention is as follows:
[0004] A graph neural network fault diagnosis method, comprising the following steps:
[0005] 1. Collect the vibration signals of bearing data, obtain samples according to the vibration signals, and perform noise reduction processing on the data through wavelet transform;
[0006] 2. Construct an adjacency matrix by calculating the Euclidean distance between nodes and construct a graph according to the k-nearest neighbor method, and divide the data into a training set and a test set;
[0007] 3. Construct a multi-modal hypersensory graph Transformer model for identifying cases where it is difficult to extract features and the extracted signals are weak under fault, noise, and vibration waveform changes. Among them, the multi-modal hypersensory graph Transformer model includes a graph data construction module, a hypersensory dilation residual module, an element-level Transformer, and a label classification module. The specific implementation method of this model is as follows:
[0008] (3a) The graph data construction module includes a fully connected layer and an adjacency matrix obtained by multiplying the feature matrix with its transpose matrix. The adjacency matrix calculates the Euclidean distance between all sample points, and then constructs a K-nearest neighbor graph based on these distances, and sets its value to 1, indicating that there is a connection edge between nodes, and the values in the remaining adjacency matrix are set to 0, indicating that there is no connection edge between nodes.
[0009] (3b) Use attribute convolution to expand the original convolution kernel of the perception field and maintain the perception with the same parameters and calculations. Different from conventional convolution, a new parameter, the dilation rate, is added to the dilation convolution, which is mainly used to increase the receptive field y of the convolution kernel. The formula is as follows:
[0010]
[0011] Where W i represents the weight of the convolution kernel; Z i·r represents the input feature map, sampled according to the dilation rate r; N represents the size of the convolution kernel. The core idea of the residual is to alleviate the vanishing gradient problem in the training of deep networks by introducing skip connections and improve the training efficiency. The main contribution of the residual is to enable the neural network to maintain good training results when the number of layers increases. The formula of the residual is as follows:
[0012]
[0013] Where Z l+1 represents the input signal matrix of the (l + 1)-th layer, H l represents the node feature matrix input in the l-th layer, A represents the normalized adjacency matrix, W l represents the weight matrix of the current layer, b l represents the bias term of the current layer, and σ represents the non-linear activation function.
[0014] (3c) The element-level Transformer is based on the Hadamard product calculation, and we normalize the score vector to highlight the important feature dimensions. The Hadamard product attention mechanism can avoid complex matrix operations, making the model structure more concise, but it can still capture the relationships between different elements in the sequence. It retains the interaction relationships between each feature dimension and can capture the influence of each feature on the final result in a fine-grained manner. The attention weights in the module and the weight matrix of the adjacency matrix The formula is as follows:
[0015]
[0016] Among them, A (h) represents the attention score matrix of the h-th attention head, M represents the adjacency matrix, and ⊙ is the Hadamard product operator. By combining the adjacency matrix to control the attention weights, the attention mechanism only propagates information between nodes connected by the graph structure; then the input features and the weight matrix of the attention weights The formula is as follows:
[0017]
[0018] Among them, X represents the input feature matrix, and W (h) represents the attention weight matrix of each head. For each head, the corresponding weight is multiplied element-wise with the input features to achieve weighted processing; the input features are weighted and transformed through the attention weights to generate independent projected feature representations for each head; then the weight matrix A of the weighted adjacency matrix weighted The formula is as follows:
[0019] A weighted = A ⊙ α
[0020] Among them, α represents the attention weight; the adjacency matrix is weighted by the attention weight to enhance the influence of important connections and at the same time suppress unimportant connections. Description of the Drawings
[0021] Figure 1 is the flowchart of the present invention;
[0022] Figure 2 is the graph of the accuracy comparison results of five random experiments;
[0023] Figure 3 is the graph of the accuracy experiment comparison results under different signal-to-noise ratios. Detailed Embodiments
[0024] The specific implementation steps of the present invention are as follows:
[0025] 1. Collect the vibration signals of bearing data, obtain samples according to the vibration signals, and preprocess the bearing vibration signals, including denoising, normalization, etc., to improve the data quality;
[0026] 2. Construct a hypersensory dilated residual connection to improve the training effect of deep networks and implement graph convolutions with different dilation rates; construct an element-level Transformer to enhance the network's learning of important features and capture the spatial dependencies and interactions between multimodal data; define an enhanced graph convolution model to implement graph convolution operations, forming a multimodal hypersensory graph Transformer model.
[0027] 3. Use the University of Paderborn bearing dataset to train and test the model, evaluate the diagnostic performance of the model, and analyze the classification performance and domain adaptation ability of the model.
[0028] The effects of the present invention are further verified by the following experiments:
[0029] The bearing dataset consists of 1,000 data samples. The sample labels of this dataset are divided into four categories: 1, 2, 3, and 4, corresponding to four types of faults that occur in fault diagnosis: artificial outer ring damage, real outer ring damage, real inner and outer ring damage, and healthy bearings. In this experiment, the vibration and torque modalities are selected for the experiment. The data format of each modality is the same. For each modality, 200 training data are randomly selected from each of the 1,000 bearing data for training, and the rest are used as test data. Each category is randomly selected and executed 5 times. Figure 2 Intuitively shows the fault diagnosis accuracy of each random experiment. Figure 3 Intuitively shows the fault diagnosis accuracy of the model under different signal-to-noise ratios; it can be seen from the experimental results that the method of the present invention still has good diagnostic accuracy in the cases of faults, noise, and changes in vibration waveforms, which also reveals that the method of the present invention is an effective bearing fault diagnosis method in the cases of faults, noise, and changes in vibration waveforms.
Claims
1. Fault diagnosis method for multi-modal hypersensitive graph Transformer, characterized in that, It includes the following steps: (1) Collect the vibration signals of bearing data, obtain samples based on the vibration signals, and perform noise reduction processing on the data through wavelet transform; (2) Construct an adjacency matrix by calculating the Euclidean distance between nodes and construct a graph according to the k-nearest neighbor method, and divide the data into a training set and a test set; (3) Construct a multi-modal hypersensory graph Transformer model for identifying faults, noise, and difficult feature extraction and weak signal extraction in the case of vibration waveform changes; among them, the hypersensory graph Transformer model includes a hypersensory dilated residual module, an element-level Transformer, a graph data construction module, and a label classification module.
2. The fault diagnosis method of the multi-modal hypersensory graph Transformer according to claim 1, wherein The specific implementation process of the hypersensory dilated residual module described in step (3) is as follows: use attribute convolution to expand the original convolution kernel of the perception field and maintain the perception with the same parameters and calculations; different from conventional convolution, a new parameter, the dilation rate, is added to the dilated convolution, which is mainly used to increase the receptive field y of the convolution kernel, and its formula is as follows: Among which W i represents the weights of the convolutional kernel, Z i·r represents the input feature map, sampled according to the dilation rate r, and N represents the size of the convolutional kernel; the core idea of the residual is to alleviate the vanishing gradient problem in the training of deep networks by introducing skip connections and improve the training efficiency; the main contribution of the residual is to enable the neural network to maintain good training results when the number of layers increases; the formula of the residual is as follows: Among which Z l+1 represents the input signal matrix of the (l + 1)-th layer, H l represents the node feature matrix of the l-th layer input, represents the normalized adjacency matrix, W l represents the weight matrix of the current layer, b l represents the bias term of the current layer, and σ represents the non-linear activation function.
3. The fault diagnosis method of the multimodal hypersensory graph Transformer according to claim 1, characterized in that The specific process of the element-level Transformer in step (3) is as follows: The element-level Transformer is based on the Hadamard product calculation, and we normalize the fractional vector to highlight the important feature dimensions; The Hadamard product attention mechanism can avoid complex matrix operations, making the model structure more concise, but it can still capture the relationships between different elements in the sequence; It retains the interaction relationships between each feature dimension and can capture the impact of each feature on the final result in a fine-grained manner; The attention weights in the module and the weight matrix of the adjacency matrix The formula is as follows: where A (h) represents the attention score matrix of the h-th attention head, M represents the adjacency matrix, and ⊙ is the Hadamard product operator; by combining the adjacency matrix to control the attention weights, the attention mechanism only propagates information between nodes connected by the graph structure; then the input features and the weight matrix of the attention weights The formula is as follows: where X represents the input feature matrix, and W (h) represents the attention weight matrix for each head; for each head, perform element-wise multiplication of the corresponding weight with the input features to achieve weighted processing; perform a weighted transformation on the input features through the attention weights to generate an independent projected feature representation for each head; then calculate the weight matrix A of the weighted adjacency matrix weighted The formula is as follows: A weighted = A ⊙ α where α represents the attention weight; the adjacency matrix is weighted by the attention weight to enhance the influence of important connections and suppress unimportant connections at the same time.
