Large bearing state identification method based on multi-sensor multi-domain feature dynamic interactive fusion

Abstract

The invention discloses a large-scale bearing state identification method based on multi-sensor multi-domain feature dynamic interactive fusion. The method comprises the steps that S1, sensors are installed at multiple positions of a large-scale bearing to collect vibration signals; s2, generating a time-frequency map, a time-frequency grey-scale map, a Gramer angle field map and a recurrence map; s3, inputting the multi-domain image data into a ResNet50 model for feature extraction; s4, by taking the multi-domain features as nodes, constructing a graph convolutional network to perform dynamic interaction of the multi-domain features; s5, generating a fused compact feature vector by using low-rank tensor fusion; s6, constructing a graph attention network, and realizing multi-sensor feature fusion through a multi-head attention mechanism; and S7, fault classification is carried out through manifold space similarity measurement, and a final diagnosis result is obtained. According to the method, dynamic interaction and fusion of multi-sensor multi-domain features are realized through the multi-stage improved graph neural network, the accuracy and credibility of the result are improved, and the technical blank of health state recognition of the large bearing is filled.

Description

Technical Field

[0001] The present invention relates to the technical field of large bearing fault monitoring and diagnosis, and particularly to a method for identifying the state of a large bearing by dynamically interacting and fusing multi-sensor multi-domain features. Background Art

[0002] As an important part of industrial machinery, the operating state of large bearings (such as the main shaft of a shield machine) directly affects the normal operation of equipment and production efficiency. Once a bearing fails, it may cause the equipment to stop, resulting in serious production losses. Therefore, it is of great significance to monitor the health status of large bearings in real time and diagnose faults.

[0003] Although there are a large number of academic papers, as well as many patents and commercial analysis software for rolling bearing fault analysis, most of these technologies and analysis methods are for ordinary small bearings at medium and high speeds. Due to the large size and numerous components of large bearings, the signal transmission path is extremely long, and it is difficult for a single sensor to comprehensively capture the vibration information of the bearing. Moreover, large bearings usually operate under low-speed and high-load conditions. Compared with high-speed roller bearings, the signal-to-noise ratio of vibration monitoring signals of low-speed roller bearings is very low, and fault signals are often submerged by strong background noise or other strong signals. The vibration characteristics in a single domain (such as the time domain or frequency domain) cannot accurately reflect the health status of the bearing. Traditional multi-sensor fusion methods are mostly used for general rolling bearings. However, for large bearings, the signals collected by sensors installed in different positions have large time delays, uneven spatial distributions, and may even be affected by different noise sources. At this time, traditional methods cannot fully model the dynamic interaction relationship between different sensor signals, resulting in low fault diagnosis accuracy. The vibration transmission path of large bearings is complex, especially in a non-linear dynamic environment (such as load changes, etc.), the vibration signal characteristics will change with time, and traditional fusion methods are difficult to adapt to this dynamic change. Summary of the Invention

[0004] In order to solve the problem that the existing technology lacks a health status recognition technology for large bearings (such as the main bearing of a shield machine and the main bearing of a wind turbine), the present invention proposes a method for identifying the state of a large bearing by dynamically interacting and fusing multi-sensor multi-domain features to solve the above problems.

[0005] The present application discloses a method for identifying the state of a large bearing by dynamically interacting and fusing multi-sensor multi-domain features, including the following steps:

[0006] S1. Install sensors at multiple positions of the large bearing to collect vibration signals;

[0007] S2. Perform sliding window sampling on the vibration signals, and generate a time-frequency diagram, a time-frequency grayscale diagram, a Gram angle field diagram, and a recurrence plot through multi-domain information transformation;

[0008] S3. Input the image data obtained in S2 into the pre-trained ResNet50 model for feature extraction to obtain the multi-domain feature vectors of each sensor;

[0009] S4. Construct a graph convolutional network, use the multi-domain feature vectors obtained in S3 as node feature vectors, generate an adjacency matrix based on the K-nearest neighbor algorithm, and extract multi-domain features through the graph convolutional network;

[0010] S5. Use low-rank tensor fusion to perform high-order interactions on the multi-domain features output by the graph convolutional network to generate the feature vectors of dynamically fused multi-domain features;

[0011] S6. Construct a graph attention network, use the multi-domain fusion feature vectors obtained for each sensor as a node, use the feature vectors after fusion in S5 as the feature vectors of the node, and obtain the feature vectors containing more hierarchical information and capable of dynamically adjusting the weights of different sensors through the multi-head attention mechanism of the graph attention network;

[0012] S7. Project the feature vectors output by the graph attention network into the manifold space, perform fault classification through the manifold space similarity metric, and obtain the final diagnosis result.

[0013] Preferably, the installation positions of the sensors include the main shaft bearing seat, the radial direction of the outer surface of the main shaft, and the axial direction of the outer surface of the main shaft.

[0014] Preferably, the transformation process of the time-frequency graph includes:

[0015] Perform empirical mode decomposition on the vibration signal to obtain multiple intrinsic mode functions;

[0016] Apply the Hilbert transform to each intrinsic mode function to obtain the instantaneous frequency and instantaneous amplitude, thereby constructing the time-frequency graph.

[0017] Preferably, the transformation process of the time-frequency grayscale graph includes:

[0018] Perform logarithmic transformation and normalization processing on the time-frequency graph, and directly map the normalized values to the pixel intensities of the grayscale graph, thereby constructing the time-frequency grayscale graph.

[0019] Preferably, the transformation process of the Gramian angular field graph includes:

[0020] Perform normalization processing on the vibration signal, then through polar coordinate transformation, convert the normalized signal into an angle representation, calculate the similarity matrix according to the angle relationship of the signal, and obtain the Gramian angular field graph of the signal.

[0021] Preferably, the transformation process of the recurrence plot includes:

[0022] Normalize the vibration signal, then perform time-delay embedding on the signal, calculate the cosine similarity of the time points in the embedding space, set a threshold and construct a similarity matrix, and use the values of the similarity matrix as pixel intensities to construct a recurrence plot.

[0023] Preferably, the generation of the adjacency matrix based on the K-nearest neighbor algorithm includes the following steps:

[0024] Calculate the cosine similarity between nodes, select the K nearest neighbor nodes for each node, and truncate the small-weight connections through a threshold.

[0025] Preferably, the low-rank tensor fusion includes the following steps:

[0026] Represent the multi-domain features as multi-modal inputs, construct a high-order tensor, and use the outer product of tensors to capture the high-order interaction relationships of different domain features;

[0027] Introduce low-rank decomposition, decompose the high-dimensional tensor into a combination of multiple low-rank tensors, and fuse the multiple low-rank tensors into a feature vector as the output of the graph convolutional network.

[0028] Preferably, the propagation process of the graph attention network is as follows:

[0029]

[0030] where, e ij is the attention coefficient, h i is the feature vector of node i, h j is the feature vector of node j, W is the shared linear transformation weight matrix, || represents the concatenation operation, LeakyReLU is the non-linear activation function, α is the learnable attention weight vector, α ij is the normalized attention coefficient between nodes under each attention head, is the normalized attention coefficient of the k-th attention head, is the set of neighbor nodes of node i, exp(·) is the exponential function, σ is the width parameter controlling the similarity range, K is the number of attention heads, W k is the weight matrix of the k-th attention head, h′ i is the feature vector after multi-head averaging.

[0031] Preferably, the calculation formula of the manifold space similarity metric is as follows:

[0032]

[0033] where, S ij is the data point similarity metric using the Gaussian kernel, ||x i -x j || 2is the Euclidean distance between data points, D ij is the manifold distance between nodes, β is a hyperparameter controlling the importance of the manifold distance, h i ″ is the final representation of node i and also the input of the classifier, serving as the result of bearing health state recognition.

[0034] Advantages of the present invention:

[0035] The invention proposes a method for large bearing state recognition based on multi-sensor multi-domain information fusion. By multi-domain information fusion and multi-level improved graph neural network feature extraction, the deep connection between bearing vibration and faults is deeply explored. Compared with traditional methods, the method proposed by the present invention can effectively improve the accuracy and credibility of the results, fill the technical gap in the health state recognition of large bearings (such as the main bearings of shield machines and wind turbines), and provide a new and effective solution for the state monitoring of large bearings. Description of the Drawings

[0036] Figure 1 is a schematic diagram of the method for large bearing state recognition with dynamic interactive fusion of multi-sensor multi-domain features according to an embodiment of the present invention;

[0037] Figure 2 is a flowchart of the method for large bearing state recognition with dynamic interactive fusion of multi-sensor multi-domain features according to an embodiment of the present invention. Detailed Embodiment

[0038] To make the objectives, technical solutions and advantages of the present application more clear, the following examples are given with reference to the accompanying drawings to further elaborate on the present application in detail.

[0039] The embodiments of the present application disclose a method for large bearing state recognition with dynamic interactive fusion of multi-sensor multi-domain features. The schematic diagram and flowchart are respectively as Figure 1 and Figure 2 shown, and include the following steps:

[0040] S1. Install multiple acceleration sensors on the main bearing seat of the large bearing, the radial direction of the outer surface of the main shaft, and the axial direction of the outer surface of the main shaft to collect bearing vibration signals at different measuring point positions.

[0041] S2. Perform sliding window sampling on the vibration signals and generate time-frequency diagrams, time-frequency grayscale diagrams, Gramian angular field diagrams, and recurrence plots through multi-domain information transformation.

[0042] Due to the relatively low rotational speed of the main bearing of the shield machine, traditional time-domain and frequency-domain signal processing methods may not be able to accurately reflect the fault vibration information. In this case, the multi-domain signal feature information combining time-frequency diagrams, time-frequency grayscale diagrams, Gramian angular field diagrams, and recurrence plots used in this embodiment can significantly improve the accuracy and robustness of the fault information. Therefore, after the original signal is sampled by a sliding window, the vibration signal of each sample is processed to obtain the above four signal images. The specific transformation process and calculation formulas are as follows:

[0043] Time-frequency diagram

[0044] Perform empirical mode decomposition (EMD) on the vibration signal to decompose the vibration signal into multiple intrinsic mode functions (IMFs). Each IMF represents a frequency band of the signal:

[0045]

[0046] where x(t) is the vibration signal, IMF i (t) is the i-th intrinsic mode function, r n (t) is the final residue, and t represents time.

[0047] Apply the Hilbert transform to each intrinsic mode function to obtain the instantaneous frequency and instantaneous amplitude, thereby constructing the time-frequency diagram:

[0048]

[0049] where is the i-th intrinsic mode function, H{·} represents the Hilbert transform, z(t) is a complex signal containing the real part (the original signal) and the imaginary part (the result of the Hilbert transform), A(t) is the instantaneous amplitude, representing the amplitude of the signal at each moment, which is the derivative of the phase of the analytic signal, f(t) is the instantaneous frequency, describing the change of the signal frequency over time, which is the modulus of the analytic signal, and TF(t,f) is the finally obtained time-frequency diagram.

[0050] Time-frequency grayscale diagram

[0051] Perform logarithmic transformation and normalization on the time-frequency diagram. Each element of the time-frequency diagram can be regarded as a pixel in the image. The normalized value is directly mapped to the pixel intensity of the grayscale diagram. The higher the value, the brighter the pixel, and the lower the value, the darker the pixel, thereby constructing the grayscale diagram. The specific calculation rules are as follows:

[0052] Logarithmic transformation: TF log (t,f) = log(1 + TF(t,f));

[0053] Normalization:

[0054] Construct the grayscale image: Gray(t,f) = round(255 × TF norm (t,f));

[0055] where, max(TF log ) is the maximum value of all elements in the time-frequency diagram TF(t,f), and min(TF l0g ) is the minimum value of all elements in the time-frequency diagram TF(t,f). After normalization, the values of the time-frequency diagram are scaled to the range [0,1]. Gray(t,f) is the pixel value at the corresponding position (t,f) in the grayscale image, with a range of [0, 255], representing the grayscale intensity of the image.

[0056] Gramian Angular Field (GAF) diagram

[0057] Normalize the vibration signal, and then through polar coordinate transformation, convert the normalized signal into an angular representation. According to the angular relationship of the signal, calculate the similarity matrix of the signal to obtain the Gramian Angular Field (GAF) diagram of the signal. The specific calculation rules are as follows:

[0058] Normalization:

[0059] Polar coordinate transformation: θ(t) = 2πx norm (t);

[0060] Similarity matrix calculation: S(t i ,t j ) = cos(θ(t i ) - θ(t j ));

[0061] where, θ(t) is the angle after the conversion of the normalized signal x norm (t), and S(t i ,t j ) represents the similarity between time points t i and t j . Based on the similarity matrix S, generate the Gramian Angular Field (GAF) diagram.

[0062] Recurrence plot

[0063] Normalize the vibration signal, and then perform time-delay embedding on the signal to convert the one-dimensional time series signal into a multi-dimensional vector. Use the signal after delay embedding to calculate the cosine similarity of the time points in the embedding space, set a threshold and construct a similarity matrix, and use the values of the similarity matrix as pixel intensities to construct a recurrence plot. The specific calculation rules are as follows:

[0064] Normalization processing:

[0065] Time-delay embedding: y(t) = [x norm(t), x norm (t + τ), …, x norm (t + (m - 1)τ)];

[0066] Cosine similarity calculation:

[0067] Recursive graph construction:

[0068] Where y(t) is the delayed embedding vector at time point t, τ is the time delay, m is the embedding dimension, ‖y(t i )‖ and ||y(t j )|| are the Euclidean norms of the vectors. Set the threshold to ∈. If S(t i , t j ) is greater than this threshold, then it is considered that there is "similarity" between t i and t j , and it is marked as "1" in the recursive graph; otherwise, it is marked as "0".

[0069] The time-frequency graph can effectively capture the time-frequency characteristics of non-stationary signals, which helps to understand the dynamic changes of the signals; the time-frequency grayscale graph enhances the feature information of the signals through normalization and grayscale mapping; the Gramian angular field graph retains the global characteristics of the time series data, which helps to capture more complex relationships in the time series, and its sparsity eliminates the redundant information between multimodals and improves the interpretability of the features; the recursive graph intuitively shows the dynamic changes of the signals, which helps to understand the non-linear characteristics of the signals.

[0070] S3. Input the image data obtained in S2 into the pre-trained ResNet50 model for feature extraction to obtain the multi-domain feature vectors of each sensor.

[0071] Normalize the generated image data to the input size and format required by the ResNet model. Use the ResNet50 model pre-trained on ImageNet to extract features from the four signal images of the time-frequency graph, time-frequency grayscale graph, Gramian angular field graph, and recursive graph obtained after transformation, and obtain the signal feature vectors of four different domains for each sensor. For the convenience of the next step of processing, normalize the obtained feature vectors. The ResNet50 network structure is shown in Table 1.

[0072] Table 1 ResNet50 network structure and parameters

[0073]

[0074] S4. Construct a graph convolutional network (GCN), use the multi-domain feature vectors obtained in S3 as the node feature vectors, generate an adjacency matrix based on the K-nearest neighbor algorithm, and extract the multi-domain features through the graph convolutional network.

[0075] Given the effectiveness of graph convolutional networks (GCNs) in processing graph data, especially their ability to aggregate the feature information of neighboring nodes through convolutional operations, thereby capturing the local structure and global information of graph data, this embodiment proposes an improved GCN network to extract multi-domain features. The signal feature vectors obtained by each subsample in S3 are used as the node feature vectors of the GCN, and the adjacency matrix is established using the K-nearest neighbor algorithm. Specifically, the cosine similarity between nodes is calculated, the K nearest neighbor nodes of each node are selected, and small-weight nodes are truncated by a threshold to generate the adjacency matrix of the graph, completing the construction of the GCN network. The GCN graph network structure is shown in Table 2:

[0076] Table 2 GCN Network Structure and Parameters

[0077]

[0078] S5. Use low-rank tensor fusion (LRTF) to perform high-order interactions on the multi-domain features output by the graph convolutional network, generating a feature vector for dynamic fusion of multi-domain features.

[0079] Since there are potential complex interactions between the features of different domains (time-frequency graphs, grayscale graphs, Gram angular field graphs, recurrence graphs), traditional pooling cannot explicitly capture these high-order interaction information. And the importance of different domain features for the fault diagnosis task may vary, and traditional pooling cannot be dynamically adjusted. To address the above deficiencies, low-rank tensor fusion is used as the pooling method for the graph network. Specifically, the features of different domains (such as time-frequency graphs, grayscale graphs, etc.) are represented as multi-modal inputs, a high-order tensor is constructed, and tensor operations (outer product) are used to capture the high-order interaction relationships between different domain features. To reduce the computational complexity of the high-order tensor, low-rank decomposition is introduced, decomposing the high-dimensional tensor into a combination of multiple low-rank tensors, and fusing multiple low-rank tensors into a feature vector as the output of the graph convolutional network. The specific calculation process is as follows:

[0080] Multi-modal feature tensoring: χ = cat(x1, x2,..., x N );

[0081] Tensor outer product:

[0082] Low-rank tensor decomposition:

[0083] where cat(·) represents stacking N feature vectors together by dimension, x i is the feature vector of the i-th domain, is the high-order vector, ⊙ is the Hadamard product of matrices, is the weight matrix for each domain, represents different domains, d i is the dimension of the i-th feature vector, ri is the output dimension after transformation, R is the rank of the decomposition used to control the dimension of the decomposition, a k , b k , c k are respectively the tensor factor matrices after low-rank decomposition, k is a positive integer from 1 to R, and ° represents the outer product of tensors.

[0084] Low-rank tensor fusion can capture the complex non-linear interactions between different domain features, retain more domain feature information, reduce redundancy through low-rank decomposition instead of directly discarding, ensure that the features still have sufficient representation ability after fusion, and can dynamically adjust the importance of each domain feature through learning, rather than simply averaging or splicing modal features. If a certain modal feature is affected by noise interference, low-rank tensor fusion can dynamically balance the weights of each domain and avoid the significant impact of single-domain noise on the diagnosis result. Through the above steps, the feature vectors after the interaction and aggregation of multi-domain vibration information of each sensor are obtained for subsequent fault diagnosis.

[0085] S6. Construct a graph attention network (GAT), using the multi-domain fusion feature vectors obtained by each sensor as a node, and the feature vectors after S5 fusion as the feature vectors of the nodes. Through the multi-head attention mechanism of the graph attention network, feature vectors containing more hierarchical information and capable of dynamically adjusting the weights of different sensors are obtained.

[0086] Large bearing structures are complex in composition and huge in volume, containing multiple interrelated components. The vibration signal sources have complex non-linear and non-stationary characteristics. The signals are often accompanied by strong noise and interference, especially under low-speed and strong non-stationary working conditions. Traditional machine learning (such as support vector machines, random forests) and neural networks (such as CNN, RNN) usually have difficulty directly dealing with the correlation and complex topological relationships of multi-domain vibration signals, resulting in reduced robustness and generalization of the model. The graph attention network (GAT) can adapt to input data with different distributions and the non-linear and high-dimensional complex vibration signal characteristics. It can effectively fuse features from different sources and capture the correlation and complementarity between multi-source features. The attention mechanism of GAT can dynamically adjust the importance weights of multi-domain features, thus more effectively capturing the structural information and feature associations in graph data, enhancing the contribution of key source features to fault diagnosis, and still having good robustness in the presence of noise or incomplete features. Therefore, GAT is constructed as the basic model for fault diagnosis. Using each sensor as a node, normalizing the feature vectors of the sensors obtained in S5 as the feature vectors of the nodes, and using the fully connected method to obtain the initial adjacency matrix, which is propagated in GAT, enabling each node to interact with its adjacent nodes under the multi-head graph attention mechanism to obtain feature vectors containing more hierarchical information. The specific propagation process of GAT is as follows:

[0087]

[0088] Among them, e ij is the attention coefficient, h i is the feature vector of node i, h j is the feature vector of node j, W is the shared linear transformation weight matrix, || represents the concatenation operation, LeakyReLU is the non-linear activation function, α is the learnable attention weight vector, α ij is the normalized attention coefficient between nodes under each attention head, is the normalized attention coefficient of the k-th attention head, is the set of neighbor nodes of node i, exp(·) is the exponential function, σ is the width parameter controlling the similarity range, K is the number of attention heads, W k is the weight matrix of the k-th attention head, h i ′ is the feature vector after multi-head averaging.

[0089] The GAT network structure is shown in Table 3:

[0090] Table 3 GAT Network Structure and Parameters

[0091]

[0092] S7. Project the feature vector output by the graph attention network into the manifold space, and perform fault classification through the manifold space similarity metric to obtain the final diagnosis result.

[0093] After node propagation in the ordinary GAT, fault classification is performed through the fully connected layer and the softmax layer. This method loses the spatial structure, temporal relationship or topological structure of the data. To solve this problem, this embodiment uses the manifold space similarity metric to retain the local structure and global structure of the data, improve the decision-making accuracy, and enhance the performance of the classifier. Through manifold learning, capture the non-linear characteristics of the data and the inherent low-dimensional structure of the high-dimensional data. By measuring the similarity between data and the distance on the manifold, identify the relationship between data and use it as the decision basis for classifier classification to obtain the final fault diagnosis result. The specific calculation rules are as follows:

[0094]

[0095] Among them, S ij is the data point similarity metric using the Gaussian kernel, ||x i -x j || 2 is the Euclidean distance between data points, D ij is the manifold distance between nodes, β is the hyperparameter controlling the importance of the manifold distance, h i″It is the final representation of node i and also the input of the softmax classifier, serving as the result of bearing health state recognition.

[0096] In summary, the embodiment of the present application discloses a large bearing state recognition method for dynamic interaction and fusion of multi-sensor multi-domain features, which is composed of two-level improved graph neural networks. The fault diagnosis method obtained through the above steps optimizes the whole process of vibration signal from acquisition to processing and then to diagnosis, can realize the dynamic interaction and fusion of multi-sensors and multi-domain features, effectively overcome the deficiencies of traditional methods, has a high accuracy and credibility, provides a new and effective solution for the health state recognition of large bearings, and can effectively identify the bearing health state under the conditions of low speed, high load and strong noise of large bearings.

[0097] The above has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features, characterized in that: The following steps are involved: S1. Install sensors at multiple locations on large bearings to collect vibration signals; S2, performing sliding window sampling on the vibration signal, and generating a time-frequency graph, a time-frequency grayscale graph, a Gram angle field graph and a recurrence graph through multi-domain information transformation; S3, input the image data obtained in S2 into the pre-trained ResNet50 model for feature extraction to obtain the multi-domain feature vector of each sensor; S4, construct a graph convolutional network, use the multi-domain feature vector obtained in S3 as the node feature vector, generate an adjacency matrix based on the K nearest neighbor algorithm, and extract multi-domain features through the graph convolutional network; S5. Use low-rank tensor fusion to perform high-order interactions on the multi-domain features output by the graph convolutional network to generate a feature vector of dynamic fusion of multi-domain features; S6. Build a graph attention network, take the multi-domain fusion feature vector obtained by each sensor as a node, take the feature vector fused by S5 as the feature vector of the node, and use the multi-head attention mechanism of the graph attention network to obtain a feature vector containing more hierarchical information and dynamically adjusting the weights of different sensors; S7. Project the feature vector output by the graph attention network into the manifold space, perform fault classification through the manifold space similarity metric, and obtain the final diagnosis result.

2. The large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features according to claim 1 is characterized in that: The installation positions of the sensor include the main shaft bearing seat, the radial direction of the main shaft outer surface and the axial direction of the main shaft outer surface.

3. The large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features according to claim 2 is characterized in that: The transformation process of the time-frequency diagram includes: Perform empirical mode decomposition on the vibration signal to obtain multiple intrinsic mode functions; The Hilbert transform is applied to each eigenmode function to obtain the instantaneous frequency and instantaneous amplitude, thus constructing the time-frequency plot.

4. The large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features according to claim 3 is characterized in that: The transformation process of the time-frequency grayscale image includes: The time-frequency graph is logarithmically transformed and normalized, and the normalized value is directly mapped to the pixel intensity of the grayscale image to construct the time-frequency grayscale image.

5. The large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features according to claim 4 is characterized in that: The transformation process of the Gram angle field diagram includes: The vibration signal is normalized, and then converted into an angle representation through polar coordinate transformation. The similarity matrix is ​​calculated according to the angle relationship of the signal to obtain the Gram angular field diagram of the signal.

6. The large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features according to claim 5 is characterized in that: The transformation process of the recursive graph includes: The vibration signal is normalized and then time-delayed and embedded. The cosine similarity of the time points in the embedding space is calculated. The threshold is set and the similarity matrix is ​​constructed. The values ​​of the similarity matrix are used as pixel intensities to construct a recursive graph.

7. The large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features according to claim 6 is characterized in that: The method for generating an adjacency matrix based on the K nearest neighbor algorithm comprises the following steps: Calculate the cosine similarity between nodes, select the K nearest neighbors of each node, and truncate small weight connections by thresholding.

8. The large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features according to claim 7 is characterized in that: The low-rank tensor fusion comprises the following steps: Represent multi-domain features as multi-modal inputs, construct high-order tensors, and use tensor outer products to capture high-order interactions between features in different domains; Low-rank decomposition is introduced to decompose the high-dimensional tensor into a combination of multiple low-rank tensors, and multiple low-rank tensors are fused into a feature vector as the output of the graph convolutional network.

9. The large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features according to claim 8 is characterized in that: The propagation process of the graph attention network is as follows: Among them, e ij is the attention coefficient, h i is the feature vector of node i, h j is the feature vector of node j, W is the shared linear transformation weight matrix, || represents the concatenation operation, LeakyReLU is the nonlinear activation function, α is the learnable attention weight vector, α ij is the normalized attention coefficient between nodes under each attention head, is the normalized attention coefficient of the kth attention head, is the set of neighbor nodes of node i, exp(·) is the exponential function, σ is the width parameter that controls the similarity range, K is the number of attention heads, and W k is the weight matrix of the kth attention head, h i ′ is the feature vector after long-term averaging.

10. The large bearing state recognition method based on dynamic interactive fusion of multi-sensor multi-domain features according to claim 9 is characterized in that: The calculation formula of manifold space similarity measure is as follows: Among them, S ij is the data point similarity measure using Gaussian kernel, ||x i -x j || 2 is the Euclidean distance between data points, D ij is the manifold distance between nodes, β is a hyperparameter that controls the importance of manifold distance, and h i ″ is the final representation of node i and is also the input of the classifier as the result of bearing health status identification.

Images