A multi-sensor fault diagnosis method based on a graph neural network

Abstract

The present application relates to the technical field of mechanical fault diagnosis and health monitoring, and particularly relates to a multi-sensor fault diagnosis method based on a graph neural network, the present application realizes a high accuracy of 95%-100% on multiple public data sets such as CWRU, PU, MFPT, SEU and the like through multi-relation fusion graph construction, node feature coding, topology-aware graph neural propagation and gate residual fusion, and is significantly superior to prior art methods in terms of weak fault recognition, the present application constructs an end-to-end framework of graph construction-feature coding-graph neural propagation-graph level reading-classification prediction through the synergistic effect of six steps, unifies the multi-stage processing mode in traditional fault diagnosis, breaks through the technical bottlenecks such as single structure modeling and one-sided feature expression in the prior art, and provides a new paradigm for intelligent fault diagnosis with high reliability and high robustness.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis and health monitoring technology, specifically to a multi-sensor fault diagnosis method based on graph neural networks. Background Technology

[0002] Fault diagnosis of rotating machinery (such as bearings, gearboxes, and motors) is crucial for ensuring equipment safety and production continuity. Multi-sensor vibration monitoring technology collects vibration signals by deploying distributed sensor networks, providing a multi-dimensional data foundation for fault identification. However, existing diagnostic technologies face three core bottlenecks: First, traditional methods such as time-frequency analysis and wavelet transform combined with SVM / random forest heavily rely on manual feature engineering, have weak generalization ability, and are difficult to uncover implicit correlations between sensors. Deep learning methods, such as 1D / 2DCNN and LSTM, although achieving end-to-end learning, treat multi-sensor data as independent channels or regular grids, severing the topological connections, energy transfer paths, and signal statistical correlations of sensors in physical space, resulting in the loss of key structural information.

[0003] Secondly, existing graph neural networks (GNNs) attempt to construct sensor graph structures, but they generally suffer from single-graph dependency defects: they only use physical topology graphs or KNN similar graphs, failing to synergistically integrate three complementary topologies: physical priors (mechanical connection relationships), data-driven (signal similarity), and local continuity (radius neighborhood). At the same time, node features are mostly limited to single representations of time-domain waveforms or spectra, lacking joint encoding of impact features (time domain) and fault frequencies (frequency domain), which weakens the ability to characterize complex faults.

[0004] In summary, there is an urgent need for a diagnostic framework that deeply integrates multi-source graph structure, node feature encoding, and topology perception mechanism to overcome technical bottlenecks such as single structural modeling and one-sided feature representation, and to provide a new paradigm for highly reliable and robust intelligent fault diagnosis. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a multi-sensor fault diagnosis method based on graph neural networks, which solves the problems of low accuracy in fault diagnosis of rotating machinery, difficulty in identifying weak faults, and lack of structural information.

[0006] To solve the above problems, the technical solution of the present invention is as follows: A multi-sensor fault diagnosis method based on graph neural networks, comprising the following steps: S1. Slice the vibration signals collected by multiple sensors into windows, and use each signal segment as a graph node to construct a multi-relationship fusion graph structure that includes a similarity graph, a radius graph, and a physical topology graph; S2. Construct node features for the vibration signal segments corresponding to the graph nodes. For the original vibration signal, you can choose to directly use the time domain signal segments as node input features to extract information reflecting impact changes and temporal dynamics; or you can perform frequency domain transformation on the signal segments to obtain frequency domain features that characterize the frequency and energy distribution of fault features. Then, input the selected features into the node encoder and map them into node embedding vectors of a unified dimension, which will be used as input for subsequent graph neural network propagation and fault classification. S3. Input the node feature vectors into the graph neural network, and perform feature propagation and enhancement through local spectral domain filtering and graph structure-constrained self-attention mechanism to obtain optimized node representations; S4. Perform gated residual fusion on the optimized node representation, adaptively balance local and global features, and then embed and aggregate all nodes into a graph-level representation vector through graph-level readout operation. S5. Based on graph-level representation vectors, the health status of equipment is predicted by a classifier to achieve fault type identification.

[0007] Furthermore, in S1, the construction of the similarity graph includes: Calculate the Euclidean distance between the feature vectors of the nodes: , in, and Representing nodes respectively and nodes eigenvectors, Represents the L2 norm; Select the nearest node for each node Edge connections are established between each neighbor to form a KNN similarity graph.

[0008] Furthermore, in S1, the construction of the radius map includes: Set distance threshold A connection is established when the distance between nodes is less than this threshold. The edge is defined as follows: , in, Represents a node With nodes The edge between, and Representing nodes respectively and nodes eigenvectors, This is the preset distance threshold.

[0009] Furthermore, in S1, the construction of the physical topology graph includes: Define the adjacency matrix based on the physical connection relationships of the mechanical equipment: , in, Represents a node With nodes The edge between, Represents a node The corresponding mechanical components, the physical connection relationship includes the structural relationship of the bearing inner ring-outer ring-rolling element or the arrangement relationship of the sensor positions on the transmission chain.

[0010] Furthermore, in S2, the frequency domain transformation includes: Perform a Fast Fourier Transform on each vibration signal segment to obtain its frequency domain features: , in, For frequency domain characteristics, Indicates the first The original vibration signal of each node, FFT Represents the Fast Fourier Transform. Represents frequency domain characteristics; In the time-domain features, the node features are directly derived from the time-domain vibration segments themselves: , Specifically, the time-domain signal is divided into impact components, vibration amplitude changes, and non-stationary dynamic trends, and a GNN model adapted to capture instantaneous changes is constructed. Logarithmic magnitude compression and Z-score normalization are used to reduce dimensional differences: , in, Indicates the first The standardized feature vector of each node's features after Z-score standardization The mean, The standard deviation is denoted as .

[0011] Furthermore, in S3, local spectral domain filtering employs Chebyshev polynomials to perform local filtering on the nodes: , in, For the normalized Laplace matrix, For Chebyshev polynomials, This represents the highest order of the Chebyshev polynomial, controlling the frequency range of the spectral domain filter. For learnable projection, The input node feature matrix, This represents the filtered node features.

[0012] Furthermore, in S3, the graph structure-constrained self-attention mechanism includes: Calculate the query matrix Key matrix Sum matrix : , in, The input node feature matrix, 、 、 These are the projection weight matrices for the query, key, and value, respectively. Define structure bias : , in, With preference for neighbor-to-neighbor propagation, These are learnable or preset parameters for structural bias. The distance threshold for structural bias is not a preset fixed constant, but is dynamically determined based on the feature distribution of the current batch of data. Specifically, this paper adopts the K-percentile strategy: first, the Euclidean distance matrix between all node pairs is calculated, and then... Set as the front of this distance distribution percentile, Representing the adjacency matrix The Element, indicating node With nodes Does an edge connect them? Represents a node With nodes The shortest path distance on the graph; The structural constraint attention score is calculated as follows: , in, Represents a node For nodes Structural constraints attention weights For nodes The query vector, For nodes The key vector, The dimension of the key vector. Indexes of other nodes used for attention summation.

[0013] Furthermore, in S4, gated residual fusion includes: Local spectral domain filtering output Self-attention output with graph structure constraints To merge: , Among them, the gating coefficient For node-level learnable gates: , This represents vector concatenation. This represents a linear transformation, with sigmoid as the activation function. This represents the final node after fusion. The total number of stacked Transformer layers in the graph; Graph-level readout operations use global average pooling: , in, Vectors are represented at the graph level. For nodes The final characteristic representation.

[0014] Furthermore, in S5, classification prediction uses a fully connected layer and a Softmax classifier, and its calculation process is as follows: , in, Vectors are represented at the graph level. and The weights and biases of the classifier. This represents the probability distribution of predicted fault types.

[0015] Compared with existing technologies, the present invention has the following advantages: by constructing a multi-relationship graph by integrating similarity graphs, radius graphs and physical topology graphs, it simultaneously characterizes the global similarity, local continuity and real physical structure of sensor signals, and completely models the spatial topology, coupling correlation and fault propagation path between sensors, thus fundamentally solving the defects of traditional deep learning models that fragment structural information and rely solely on Euclidean space learning.

[0016] By employing a method of jointly extracting time-domain impulse features and frequency-domain fault features, and fusing them with dual-channel coding, along with logarithmic amplitude compression and Z-score standardization, the problem of large differences in signal amplitude and strong noise interference under different sensors and operating conditions is effectively solved. At the same time, the frequency of time-domain impulse and frequency-domain fault features is preserved, which significantly improves the identification accuracy of early weak faults and small sample faults.

[0017] By combining local spectral domain filtering with graph structure-constrained self-attention, we can accurately extract local neighborhood fault features and suppress noise, as well as model long-range dependencies across sensors. With the help of a gated residual fusion mechanism, we can adaptively balance local and global features, effectively solving the problems of excessive smoothing and insufficient expressive power in deep propagation of traditional GNNs. The model can be extended to deeper levels.

[0018] The multi-relationship graph structure and node feature encoding make the model highly robust to changes in rotational speed, load fluctuations, and noise interference. The diagnostic accuracy fluctuates very little under various operating conditions. The unified end-to-end framework can be directly transferred to different rotating machinery and different sensor deployment scenarios without significant parameter adjustments, and it has strong industrial generalization. Attached Figure Description

[0019] Figure 1This is a flowchart illustrating the overall process of multi-sensor fault diagnosis in this invention. Figure 2 This is a schematic diagram illustrating the construction of the multi-relationship graph of the present invention; Figure 3 This invention provides a graph neural network fault diagnosis framework for the TST-GT model. Figure 4 This is a flowchart of the time-frequency feature encoding process of the present invention; Figure 5 Results of fault diagnosis tasks on four different datasets; Figure 6 This is a comparison chart of fault diagnosis accuracy across multiple datasets. Detailed Implementation

[0020] This invention presents a multi-sensor fault diagnosis method based on graph neural networks. Addressing the shortcomings of existing multi-sensor fault diagnosis technologies, such as lack of structural modeling, simplistic graph construction, insufficient node feature representation, and the tendency of traditional graph neural networks to oversmooth, this invention achieves end-to-end, high-precision, and robust rotating machinery fault identification through multi-relation fusion graph construction, node feature encoding, topology-aware graph neural propagation, and gated residual fusion. The diagnostic method is as follows: Figure 1 As shown, the steps are as follows: S1. Slice the vibration signals collected by multiple sensors into windows, and use each signal segment as a graph node to construct a multi-relationship fusion graph structure that includes a similarity graph, a radius graph, and a physical topology graph; S2. Construct feature representations for the vibration signal segments corresponding to the graph nodes; where the node features adopt a single-path input method: directly select the time-domain vibration segment as the node's time-domain feature, or select the frequency-domain spectral feature as the node's frequency-domain feature after frequency-domain transformation of the vibration signal; then input the selected node features into the encoder, mapping them into node embedding vectors of a uniform dimension. S3. Input the node feature vectors into the graph neural network, and perform feature propagation and enhancement through local spectral domain filtering and graph structure-constrained self-attention mechanism to obtain optimized node representations; S4. Perform gated residual fusion on the optimized node representation, adaptively balance local and global features, and then embed and aggregate all nodes into a graph-level representation vector through graph-level readout operation. S5. Based on graph-level representation vectors, the health status of equipment is predicted by a classifier to achieve fault type identification.

[0021] The specific steps and results are as follows: S1. The vibration signals acquired by multiple sensors are sliced ​​into windows, and each signal segment is used as a graph node to construct a multi-relationship fusion graph structure containing a similarity graph, a radius graph, and a physical topology graph, such as... Figure 2 As shown, the specific steps are as follows: S11. The raw vibration signals from multiple sensors are sliced ​​into fixed time windows. Each slice is regarded as a node in the graph. The window length is determined according to the sampling frequency and fault characteristic bandwidth. The windows can be set to overlap to enhance data density.

[0022] S12, KNN similarity graph construction: Calculate the Euclidean distance in the feature space for all nodes, given the feature vector of each node. Nodes are calculated using Euclidean distance. With nodes Distance metric between: ,in, and Representing nodes respectively and nodes eigenvectors, Represents the L2 norm; Select the nearest node for each node By establishing edge connections between each neighbor, a KNN similarity graph is formed. The KNN similarity graph can capture global similarity patterns between data points. Even if they are not physically adjacent, they can find neighbors in the data space, thereby uncovering potential fault associations.

[0023] S13. Constructing a radius map and setting distance thresholds. A connection is established when the distance between nodes is less than this threshold. The edge is defined as follows: , in, Represents a node With nodes The edge between, and Representing nodes respectively and nodes eigenvectors, This is the preset distance threshold.

[0024] Radius graphs ensure that each node has a certain number of local neighbors, enhancing the local continuity of the graph, avoiding the island problem caused by sparse nodes in certain regions of the KNN graph, and improving the connectivity and stability of the graph.

[0025] S14. Physical Topology Construction: Based on the mechanical equipment structure, sensor placement, and energy transfer paths, a fixed topology is established. The adjacency matrix of the physical topology is defined by the following formula: , in, Represents a node The corresponding mechanical components, the physical connection relationship includes the structural relationship of the bearing inner ring-outer ring-rolling element or the arrangement relationship of the sensor positions on the transmission chain.

[0026] S2. Construct node features for the vibration signal segments corresponding to the graph nodes. For the original vibration signal, the time-domain signal segment can be directly used as the node input feature to extract information reflecting impact changes and temporal dynamics; or the signal segment can be frequency-domain transformed to obtain frequency-domain features characterizing the frequency and energy distribution of fault features. Subsequently, the selected features are input into the node encoder and mapped to a node embedding vector of uniform dimension, which serves as the input for subsequent graph neural network propagation and fault classification. Figure 4 As shown, the specific steps are as follows: S21. In the time-domain features, the nodal features are directly derived from the time-domain vibration segments themselves: , The time-domain signal is divided into impact components, vibration amplitude changes, and non-stationary dynamic trends. In this way, a GNN model adapted to capture instantaneous changes is constructed. S22. Frequency domain feature extraction: Perform Fast Fourier Transform (FFT) on each segment to obtain the amplitude spectrum. , Logarithmic magnitude compression and Z-score normalization are used to reduce dimensional differences and improve feature separability. , S23. Unified encoder generates node embeddings: A lightweight MLP maps time-domain and frequency-domain features to obtain initial node embedding vectors. , , in, For nodes The initial feature embedding, This represents the encoder mapping function, used to map raw features to the node embedding space. , For learnable weights, , For bias, This is the ReLU activation function.

[0027] S3. Input the node feature vectors into the graph neural network, and perform feature propagation and enhancement through local spectral domain filtering and graph structure-constrained self-attention mechanism to obtain optimized node representations, such as... Figure 3 As shown, the specific steps are as follows: S31. Local spectral domain filtering: Spectral domain filtering can smooth the graph signal, suppress high-frequency noise, and retain low-frequency, meaningful fault features. It is equivalent to performing a low-pass filter in the frequency domain of the graph. The steps are as follows: Local filtering of nodes using Chebyshev polynomials: , in, For the normalized Laplace matrix, For Chebyshev polynomials, For learnable projection, The input node feature matrix, This represents the filtered node features.

[0028] To simplify the implementation, it is equivalent to using two layers of GCN: , in, Denotes the adjacency matrix with added self-loops, where It is the identity matrix. The degree matrix after adding self-loops, , and This is the learnable weight matrix of the GCN layer.

[0029] S32. Graph structure-constrained self-attention mechanism: By introducing structural bias, the attention mechanism is guided to focus on physically or data-related neighbors, avoiding the waste of computational resources between unconnected nodes in traditional GAT, while enhancing the model's ability to perceive graph structure. The steps are as follows: Calculate the query matrix Key matrix Sum matrix : , in, The input node is represented (at the first level as) ), , , These are the projection weight matrices for the query, key, and value, respectively.

[0030] Define structure bias : For each pair of nodes ,definition , Alternatively, learnable relative distance encoding can be used. , usually set It propagates within the preferred neighborhood.

[0031] The structural constraint attention score is calculated as follows: , Use multi-head attention ( heads), intra-layer residuals and LayerNorm: , in, This represents the forward propagation function of the graph attention Transformer layer. The representation layer normalization operation is used to stabilize the training process. This represents a random deactivation regularization operation to prevent overfitting. This represents a multi-head attention mechanism and parallel computation. Group attention and splice, This is followed by the feedforward network (FFN) and the second residual: , in, This represents the forward propagation function of the feedforward network block. A feedforward network typically consists of two linear transformations and activation functions. Multi-layer stacking (Typically 2-4 layers) can learn higher-order semantics.

[0032] S4. Gated residual fusion is applied to the optimized node representations to adaptively balance local and global features. Then, graph-level readout operations are used to embed and aggregate all nodes into a graph-level representation vector. The gated residual fusion mechanism allows the model to adaptively choose whether to retain information from the previous layer (local features) or adopt newly aggregated information (global features), effectively balancing local details and global structure, and alleviating the gradient vanishing and oversmoothing problems of deep networks. The steps are as follows: S41, Gated Residual Fusion: This involves filtering the local spectral domain output... Self-attention output with graph structure constraints To merge: , Among them, the gating coefficient For node-level learnable gates: , in, This represents vector concatenation. This represents a linear transformation, with sigmoid as the activation function. This represents the final node after merging.

[0033] S42, Graph-level readout operation: Global average pooling is used to embed and aggregate all nodes into a graph-level representation vector: , in, Vectors are represented at the graph level. For nodes The final characteristic representation.

[0034] S5. Based on graph-level representation vectors, predict the health status of equipment using a classifier to achieve fault type identification. The steps are as follows: S51. Classification Prediction: Fault type identification is achieved through a fully connected layer and a Softmax classifier. , The classification output can include various mechanical faults such as normal status, outer ring fault, inner ring fault, rolling element fault, and gear damage.

[0035] S52, Training and Optimization: Use the cross-entropy loss function for multi-class classification training, and improve stability by averaging multiple trials. , in, This represents the cross-entropy loss function value, used to measure the difference between the predicted probability and the true label. For the first The true one-hot vector of each sample's label. For the first The predicted probability distribution vector of each sample. This represents the total number of fault categories. This represents the number of training samples.

[0036] S6. Diagnostic abnormalities and robust control mechanisms, the steps are as follows: S61. Noise interference processing: Based on radius map and spectral domain filtering, feature smoothing is achieved, high-frequency noise interference is automatically suppressed, and the ability to identify weak faults is improved.

[0037] S62. Adaptive to changing operating conditions: The physical topology and frequency domain features enable the model to maintain structural invariance under different speeds, loads, and sampling conditions. All node features are aggregated into a fixed-dimensional graph representation. , in, This represents a fixed-dimensional graph representation vector after graph-level readout, used for adaptive adaptation to changing operating conditions. S63. Node feature anomaly detection: Predicting device health status through a fully connected layer and a Softmax classifier. , This is the weight matrix of the node feature anomaly detection classifier. To represent dimensions in a graph, This is the bias vector of the node feature anomaly detection classifier. If a node's features deviate from the overall distribution, the model can reduce its weight through an attention descent mechanism to avoid affecting the overall diagnosis.

[0038] S64. Automatic correction of structural conflicts: When a certain type of topology conflicts with other evidence during graph construction, such as when a radius graph ignores remote strongly related nodes, the attention mechanism automatically increases the weight of other topological paths to achieve topology adaptation.

[0039] To verify the effectiveness of the multi-sensor fault diagnosis method based on graph neural network (TST-GT) described in this invention, it was compared with various existing technical methods, including traditional machine learning methods and existing graph neural network methods.

[0040] Comparative Example 1: Traditional Machine Learning Methods MLP (Multilayer Perceptron): A three-layer fully connected neural network is used to directly concatenate vibration signals from multiple sensors into a vector input without graph structure modeling. The experimental setup is a hidden layer dimension of 128, using ReLU activation function and Dropout regularization.

[0041] Comparative Example 2: Existing Graph Neural Network Methods ChebyNet: Employs a Chebyshev multinomial spectral graph convolutional network, using only a single graph structure (KNNGraph / RadiusGraph / PhysicalGraph) with order K=3.

[0042] GCN (Graph Convolutional Network): It uses a first-order spectral graph convolution approximation, employs only a single graph structure, and has 2 layers.

[0043] SGCN (Simple Graph Convolutional Network): A simplified graph convolutional network that removes non-linear activations and uses only a single graph structure.

[0044] GraphSage: A graph neural network that employs a sampling aggregation strategy, using mean aggregation as the aggregation function and a single graph structure.

[0045] GIN (Graph Isomorphic Network): Employs an injectable graph neural network, uses an MLP aggregation function, and uses only a single graph structure.

[0046] GAT (Graph Attention Network): A graph neural network that uses a self-attention mechanism, employs only a single graph structure, and has 8 attention heads.

[0047] Comparative Example 3: The Method of the Invention (TST-GT) The invention employs a comprehensive framework for constructing a multi-relationship fusion graph (KNN similarity graph + radius graph + physical topology graph), node feature encoding, local spectral domain filtering and graph structure-constrained self-attention mechanism, and gated residual fusion.

[0048] Dataset: CWRU: Case Western Reserve University bearing dataset, which includes four states: normal, inner ring failure, outer ring failure, and rolling element failure. PU: Paderborn University Bearing Dataset, containing various fault types and different operating conditions; MFPT: Mechanical Failure Prevention Technology Dataset, containing weak faults, small sample size, and high noise data; SEU: Southeast University Gearbox Dataset, containing gear faults under different load and speed conditions.

[0049] Evaluation metrics: diagnostic accuracy (%), with performance tested under time-domain input (TD), frequency-domain input (FD), and physical graph conditions.

[0050] Node-level fault diagnosis results, such as Figure 5 As shown in the figure, TD represents the time domain input, FD represents the frequency domain input, and PhysicalGraph represents the physical topology graph.

[0051] 2. Comparison of fault diagnosis accuracy across multiple datasets, such as... Figure 6 As shown, the diagnostic accuracy of each method is compared on four datasets under the FD+PhysicalGraph condition.

[0052] Results analysis: 1. Advantages of constructing multiple relationship graphs Comparing the results of single graph structures (KNNGraph / RadiusGraph) and physical topology graphs (PhysicalGraph), it can be seen that physical topology graphs exhibit better or comparable performance on various datasets. This invention, by fusing three graph structures, can simultaneously capture signal similarity, local continuity, and physical structure relationships, which is significantly better than single graph construction methods.

[0053] 2. Advantages of time-frequency coding Comparing the results of TD (time domain) and FD (frequency domain) inputs, frequency domain input (FD) generally outperforms time domain input (TD) across all methods, especially on the PU and MFPT datasets. This invention employs node feature encoding, fully utilizing the discriminative power of frequency domain fault characteristic frequencies while preserving time domain impulse information, thus achieving joint identification of impulse-type faults and frequency modulation-type faults.

[0054] 3. Advantages of Structural Constraints on Attention Comparing GAT with the present invention TST-GT, under the PhysicalGraph condition, TST-GT exhibits different characteristics on the MFPT (70.67 vs 66.31) and SEU (50.29 vs 62.15) datasets. Overall, TST-GT enables the attention mechanism to focus on nodes related to the mechanical structure by introducing structural bias.

[0055] 4. Weak fault identification capability On the MFPT dataset with weak faults, small sample size, and high noise, the TST-GT of this invention achieves an accuracy of 96.82% under the PhysicalGraph+FD condition, which is significantly better than methods such as MLP (91.61%), GCN (94.94%), and SGCN (90.27%), verifying the ability of this invention to identify early rolling element damage.

[0056] The comparative experimental results above show that the multi-sensor fault diagnosis method based on graph neural network (TST-GT) described in this invention achieves a high accuracy of 95%-100% on multiple public datasets such as CWRU, PU, ​​MFPT, and SEU through multi-relation fusion graph construction, node feature encoding, topology sensing graph neural propagation and gated residual fusion. In particular, it is significantly superior to existing methods in weak fault identification.

[0057] This invention constructs a complete end-to-end framework of graph construction, feature encoding, graph neural propagation, graph-level readout, and classification prediction through the synergistic effect of six steps. It unifies the multi-stage processing method in traditional fault diagnosis and breaks through the technical bottlenecks of single structural modeling and one-sided feature expression in existing technologies, providing a new paradigm for highly reliable and robust intelligent fault diagnosis.

[0058] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-sensor fault diagnosis method based on graph neural networks, characterized in that: Includes the following steps: S1. Slice the vibration signals collected by multiple sensors into windows, and use each signal segment as a graph node to construct a multi-relationship fusion graph structure that includes a similarity graph, a radius graph, and a physical topology graph; S2. Construct node features for the vibration signal segments corresponding to the graph nodes. For the original vibration signal, the time domain signal segment can be directly used as the node input feature to extract information reflecting the impact change and temporal dynamics; or the signal segment can be transformed in the frequency domain to obtain the frequency domain features that characterize the frequency and energy distribution of the fault features. Then, the selected features are input into the node encoder and mapped into a node embedding vector of a unified dimension, which is used as the input for subsequent graph neural network propagation and fault classification. S3. Input the node feature vectors into the graph neural network, and perform feature propagation and enhancement through local spectral domain filtering and graph structure-constrained self-attention mechanism to obtain optimized node representations; S4. Perform gated residual fusion on the optimized node representation, adaptively balance local and global features, and then embed and aggregate all nodes into a graph-level representation vector through graph-level readout operation. S5. Based on graph-level representation vectors, the health status of equipment is predicted by a classifier to achieve fault type identification.

2. The multi-sensor fault diagnosis method according to claim 1, characterized in that, In S1, the construction of the similarity graph includes: Calculate the Euclidean distance between the feature vectors of the nodes: , in, and Representing nodes respectively and nodes eigenvectors, Represents the L2 norm; Select the nearest node for each node Edge connections are established between each neighbor to form a KNN similarity graph.

3. The multi-sensor fault diagnosis method according to claim 1, characterized in that, In S1, the construction of the radius map includes: Set distance threshold A connection is established when the distance between nodes is less than this threshold. The edge is defined as follows: ， in, Represents a node With nodes The edge between, This is the preset distance threshold.

4. The multi-sensor fault diagnosis method according to claim 1, characterized in that, In S1, the construction of the physical topology graph includes: Define the adjacency matrix based on the physical connection relationships of the mechanical equipment: ， in, Represents a node The corresponding mechanical components, the physical connection relationship includes the structural relationship of the bearing inner ring-outer ring-rolling element or the arrangement relationship of the sensor positions on the transmission chain.

5. The multi-sensor fault diagnosis method according to claim 1, characterized in that, In S2, the frequency domain transformation includes: Perform a Fast Fourier Transform on each vibration signal segment to obtain its frequency domain features: , in, For frequency domain characteristics, Indicates the first The original vibration signal of each node, FFT Represents the Fast Fourier Transform. Represents frequency domain characteristics; In the time-domain features, the node features are directly derived from the time-domain vibration segments themselves: , Specifically, the time-domain signal is divided into impact components, vibration amplitude changes, and non-stationary dynamic trends, and a GNN model adapted to capture instantaneous changes is constructed. Logarithmic magnitude compression and Z-score normalization are used to reduce dimensional differences: , in, Indicates the first The standardized feature vector of each node's features after Z-score standardization The mean, The standard deviation is denoted as .

6. The multi-sensor fault diagnosis method according to claim 1, characterized in that, In S3, local spectral domain filtering uses Chebyshev polynomials to perform local filtering on nodes: , in, For the normalized Laplace matrix, For Chebyshev polynomials, This represents the highest order of the Chebyshev polynomial, controlling the frequency range of the spectral domain filter. For learnable projection, The input node feature matrix, This represents the filtered node features.

7. The multi-sensor fault diagnosis method according to claim 6, characterized in that, In S3, the graph structure-constrained self-attention mechanism includes: Calculate the query matrix Key matrix Sum matrix : , in, The input node feature matrix, 、 、 These are the projection weight matrices for the query, key, and value, respectively. Define structural bias : , in, With preference for neighbor-to-neighbor propagation, These are learnable or preset parameters for structural bias. The distance threshold for structural bias is not a preset fixed constant, but is dynamically determined based on the feature distribution of the current batch of data. Specifically, this paper adopts the K-percentile strategy: first, the Euclidean distance matrix between all node pairs is calculated, and then... Set as the front of this distance distribution percentile, Representing the adjacency matrix The Element, indicating node With nodes Does an edge connect them? Represents a node With nodes The shortest path distance on the graph; The structural constraint attention score is calculated as follows: , in, Represents a node For nodes Structural constraints attention weights For nodes The query vector, For nodes The key vector, The dimension of the key vector. Indexes of other nodes for attention summation.

8. The multi-sensor fault diagnosis method according to claim 1, characterized in that, In S4, gated residual fusion includes: Local spectral domain filtering output Self-attention output with graph structure constraints To merge: ， Among them, the gating coefficient For node-level learnable gates: , This represents a linear transformation, with sigmoid as the activation function. This represents the final node after fusion. The total number of stacked Transformer layers in the graph; Graph-level readout operations use global average pooling: , in, Vectors are represented at the graph level. For nodes The final characteristic representation.

9. The multi-sensor fault diagnosis method according to claim 1, characterized in that, In S5, classification prediction uses a fully connected layer and a Softmax classifier, and its calculation process is as follows: , in, Vectors are represented at the graph level. and The weights and biases of the classifier. This represents the probability distribution of predicted fault types.

Images