Transmission system variable speed fault diagnosis method based on interpretable diagram wavelet activation network

Abstract

The invention discloses a transmission system variable-speed fault diagnosis method based on an interpretable diagram wavelet activation network, and the method comprises the steps: collecting vibration signals corresponding to different states of a transmission system under a variable-speed working condition, and obtaining a sample set; constructing a graph structure for the sample sets in different states; constructing an interpretable diagram wavelet activation network; the interpretable graph wavelet activation network comprises a wavelet attention layer, a stacked graph wavelet activation convolution layer and Readout operation; processing each node in each graph structure by a wavelet attention layer to obtain an enhanced low-frequency component; taking a graph structure formed by the enhanced low-frequency component and the adjacent matrix as input of a graph wavelet activation convolution layer, and obtaining feature representation after node updating; using Readout operation to fold the updated node feature representation into graph representation; and inputting the graph representation into a full connection layer to realize state identification of the transmission system under the condition of variable rotating speed. According to the method, the subsamples of signal division are constructed into the graph data, and the wavelet attention layer and the graph wavelet activation convolution layer are constructed, so that high-precision and interpretable fault diagnosis of the transmission system in a non-stationary environment is realized.

Description

Technical Field

[0001] This invention relates to a method for diagnosing variable speed faults in transmission systems based on interpretable graph wavelet activation networks, belonging to the field of mechanical fault diagnosis. Background Technology

[0002] In recent years, mechanical fault diagnosis technology has been gradually evolving from traditional signal processing and machine learning methods towards intelligent diagnosis based on deep learning. Deep learning models possess end-to-end feature extraction capabilities, enabling them to automatically learn fault-related discriminative features from large-scale vibration signals and exhibiting good diagnostic performance under conditions of sufficient data. However, these methods still face two challenges in practical engineering applications: First, deep model structures are typically complex, making it difficult to establish a clear correspondence between the feature extraction process and the physical mechanism of the signal, resulting in insufficient model interpretability and hindering fault cause tracing and reliable deployment in industrial scenarios. Second, transmission systems often operate under non-stationary conditions such as variable speeds, and vibration signals exhibit significant nonlinear and time-varying characteristics. Traditional deep models have limited performance in cross-condition transfer and robustness, which can easily lead to a decline in diagnostic performance.

[0003] To enhance the structural and adaptive nature of feature representation, graph neural networks (Graph Neural Networks) are increasingly being introduced into mechanical fault diagnosis tasks. Graph Neural Networks can construct nodes from local segments of vibration signals, forming a graph structure by modeling the similarity relationships between nodes. Based on message passing mechanisms, they enable the interaction and aggregation of neighborhood information, thereby mining cross-sample correlation structural features. This makes them more suitable for processing vibration signals with correlated patterns than traditional deep learning models. However, the aggregators and updaters of existing graph neural networks have poor interpretability, and the extracted vibration signal features may lose their physical meaning after relevant convolutions. Therefore, it is necessary to construct a graph convolutional network architecture that remains robust under varying operating conditions and whose feature extraction process is physically interpretable, in order to achieve reliable identification and diagnosis of transmission system fault states. Summary of the Invention

[0004] This invention provides a method for fault diagnosis of transmission systems with variable speed based on interpretable graph wavelet activation networks. By constructing sub-samples of the signal as graph data and constructing wavelet attention layers and graph wavelet activation convolutional layers, high-precision and interpretable fault diagnosis of transmission systems under non-stationary environments can be achieved.

[0005] The technical solution of this invention is:

[0006] According to a first aspect of the present invention, a method for diagnosing variable speed faults in a transmission system based on an interpretable graph wavelet activation network is provided, comprising:

[0007] Step 1: Collect vibration signals of the transmission system under different states during variable speed operation; normalize each vibration signal, divide the normalized signal into sub-samples of length L, and assign a corresponding state label to each sub-sample of each vibration signal. The resulting sample set is represented as follows: The status includes healthy status and fault status. This represents the i-th subsample of the j-th vibration signal. Then it means The corresponding status label, This represents the number of subsamples of the j-th vibration signal.

[0008] Step 2: Construct a graph structure for the sample sets under different states.

[0009] Step 3: Construct an interpretable graph wavelet activation network. This network includes a wavelet attention layer, stacked graph wavelet activation convolutional layers, and a Readout operation. Each node in each graph structure is processed through the wavelet attention layer to obtain enhanced low-frequency components. The enhanced low-frequency components and the graph structure formed by the adjacency matrix are used as input to the graph wavelet activation convolutional layers to obtain the updated feature representation H of the nodes. The Readout operation is then used to fold the updated node feature representation H into a graph representation. .

[0010] Step 4, Represent the diagram Input is fed into the fully connected layer to enable state recognition of the transmission system under variable speed conditions.

[0011] Furthermore, the step of dividing the normalized signal into subsamples of length L specifically involves using a sliding window of preset length to divide the normalized signal into subsamples without overlap.

[0012] Furthermore, the construction of the graph structure for the sample set under different states specifically involves: starting from the sample set... Each time, m subsamples are selected as the node set of a single image. The RadiusGraph method is used to calculate the cosine similarity between nodes. If the cosine similarity between two nodes is greater than a threshold If an edge connects two nodes, then the adjacency matrix of the graph structure can be obtained. ;p,q=1,2,...,m.

[0013] Furthermore, the wavelet attention layer specifically comprises:

[0014] For each node in each graph structure Discrete wavelet transform is used for multi-scale decomposition to obtain the low-frequency components. With high frequency components ;

[0015] Based on each node The low-frequency and high-frequency components are based on learnable attention weights. The low-frequency components are enhanced to obtain the enhanced low-frequency components:

[0016] ;

[0017] In the formula, It is a node Low-frequency components enhanced with attention.

[0018] Furthermore, the stacked graph wavelet activation convolutional layer comprises two graph wavelet activation convolutional layers, which will activate the low-frequency components of each node after attention enhancement. Adjacency matrix of graph structure The graph structure formed As input to the stacked graph wavelet activation convolutional layers, each layer performs feature activation and message passing on each node using the WKA operator constructed from wavelet basis functions with learnable scaling factor a and translation factor b, thereby obtaining the updated feature representation of the node. The feature update process of each node in the graph wavelet activation convolutional layer can be defined as follows:

[0019] ;

[0020] in, Indicates the first WKA operator for layer GWAConv; Representative node In the Feature vectors of layer-map wavelet-activated convolutional layers; For nodes The feature representation updated after the k-th layer of graph wavelet activation convolutional layer; Represents a node All adjacent nodes The set, i.e. ; Represents a node In the Feature vectors of layer-map wavelet-activated convolutional layers; These are learnable equilibrium parameters.

[0021] Furthermore, the diagram representation The input is fed into the fully connected layer to achieve state recognition of the transmission system under variable speed conditions. The expression is:

[0022] ;

[0023] in, Indicates a fully connected layer. It is a non-linear activation function. Used to map the output of the fault classifier to the probability of a state. .

[0024] According to a second aspect of the present invention, a transmission system variable speed fault diagnosis system based on an interpretable graph wavelet activation network is provided, comprising modules of any of the methods described above.

[0025] According to a third aspect of the present invention, a terminal is provided, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to perform the steps of the method described in any one of the preceding inventions.

[0026] The beneficial effects of this invention are as follows: By constructing vibration signals as graph data with internal correlation structures and introducing an interpretable graph wavelet activation network based on WKA layers, the aggregation and updating process of node features in the graph structure has clear physical interpretability; at the same time, by performing wavelet decomposition on node signals and using learnable attention weights to strengthen the effective low-frequency features of the signals, the stability and cross-condition robustness of feature representation are improved. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention.

[0028] Figure 2 This is a schematic diagram of the bearing transmission system fault simulation test bench in this embodiment.

[0029] Figure 3 This is a schematic diagram of the vibration signal sample division and correlation diagram construction process in this embodiment.

[0030] Figure 4 This is a schematic diagram of the wavelet attention layer in this embodiment.

[0031] Figure 5 This is a schematic diagram of a graph wavelet-activated convolutional layer in an embodiment of the present invention.

[0032] Figure 6 This is a schematic diagram of the fault diagnosis framework for interpretable graph wavelet activation networks in this embodiment.

[0033] Figure 7 This is a schematic diagram of the diagnostic results of the bearing transmission system fault experiment in this embodiment.

[0034] Figure 8 This is a schematic diagram illustrating the interpretability of the graph wavelet activation network in this embodiment. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0036] Example 1: As Figures 1-8 As shown, according to a first aspect of the present invention, a method for diagnosing variable speed faults in a transmission system based on an interpretable graph wavelet activation network is provided, comprising:

[0037] Step 1: Collect vibration signals of the transmission system under different states during variable speed operation; normalize each vibration signal, divide the normalized signal into sub-samples of length L, and assign a corresponding state label to each sub-sample of each vibration signal. The resulting sample set is represented as follows: The status includes healthy status and fault status. This represents the i-th subsample of the j-th vibration signal. Then it means The corresponding status label, This represents the number of subsamples for the j-th vibration signal. It should be noted that the length of the subsamples for each vibration signal is the same, and the number of subsamples may be the same or different.

[0038] Step two involves constructing a graph structure for the sample sets under different states, forming a graph structure dataset. The construction of the graph structure for each state specifically involves: starting from the sample set... Each time, m subsamples are selected as the node set of a single image. The RadiusGraph method is used to calculate the cosine similarity between nodes. If the cosine similarity between two nodes is greater than a threshold If an edge connects two nodes, then the adjacency matrix of the graph structure can be obtained. ;

[0039] Right now:

[0040] ;

[0041] In the formula, The dimension is ; The element in the p-th row and q-th column of the adjacency matrix is ​​used to determine the node. With nodes Whether there is an edge connecting them is indicated by a value of 1 (1 indicates an edge connection) and a value of 0 (0 indicates no edge connection); p,q=1,2,...,m; Let m represent the m-th node, and each node corresponds to a feature vector.

[0042] Step 3: Construct an interpretable graph wavelet activation network (GWAN). The GWAN includes a wavelet attention layer, stacked graph wavelet activation convolutional (GWAConv) layers, and a Readout operation. Each node in each graph structure is processed through the wavelet attention layer to obtain enhanced low-frequency components. The enhanced low-frequency components and the graph structure formed by the adjacency matrix are used as input to the graph wavelet activation convolutional layers to obtain the updated feature representation H of the nodes. The Readout operation is used to fold the updated node feature representation H into a graph representation. .

[0043] The wavelet attention layer is specifically as follows:

[0044] For each node in each graph structure Discrete wavelet transform is used for multi-scale decomposition to obtain the low-frequency components. With high frequency components The process can be represented as follows:

[0045] ;

[0046] In the formula, Represents a node The Amplitude information node corresponding to each sampling point The data length is L, therefore ; and These represent low-frequency filters and high-frequency filters, respectively. and Representing nodes respectively The amplitudes of the low-frequency and high-frequency components at the nth output sampling point after discrete wavelet transform decomposition; Indicates downsampling, such that The sequence length of the low-frequency and high-frequency components is approximately half the original length, meaning that the range of values ​​for n is: .

[0047] Subsequently, based on each node The low-frequency and high-frequency components are based on learnable attention weights. The low-frequency components are enhanced to obtain the enhanced low-frequency components:

[0048] ;

[0049] In the formula, It is a node Low-frequency components enhanced with attention.

[0050] The stacked graph wavelet activated convolutional layer specifically involves: [processing] the low-frequency components of each node after attention enhancement. and adjacency matrix The graph structure formed As input to the graph wavelet activation convolutional layer, each GWAConv layer performs feature activation and message passing on each node using the WKA operator (wavelet Kolmogorov-Arnold, WKA) constructed from wavelet basis functions with learnable scaling factor a and translation factor b, thereby obtaining the updated feature representation of the node. The feature update process for each node in the GWAConv layer can be defined as follows:

[0051] ;

[0052] in, Indicates the first WKA operator for layer GWAConv; Representative node In the The feature vector of layer GWAConv initially contains the features of the input nodes, i.e.: ; For nodes Feature representation updated by GWAConv at layer k; Represents a node All adjacent nodes The set, i.e. ; Represents a node In the Feature vectors of layer GWAConv; This is a learnable balancing parameter used to adjust the weight of a node's own features during the update process.

[0053] The interpretable graph wavelet activation network constructed by the method of this invention consists of two stacked GWAConv layers, which progressively aggregate and update the neighborhood information of node features. The extracted node feature representation H can be defined as:

[0054] .

[0055] Step four, The input to the fully connected layer to achieve state recognition of the transmission system under variable speed conditions can be represented as follows:

[0056] ;

[0057] ;

[0058] in, Indicates a fully connected layer. It is a non-linear activation function. Used to map the output of the fault classifier to the probability of a state. .

[0059] According to a second aspect of the present invention, a transmission system variable speed fault diagnosis system based on an interpretable graph wavelet activation network is provided, comprising modules of any of the methods described above.

[0060] According to a third aspect of the present invention, a terminal is provided, including a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to perform the steps of the method described in any one of the above embodiments.

[0061] Example 2: As Figures 1-8 As shown, in this embodiment, the bearing dataset from Huazhong University of Science and Technology (HUST) is used. Figure 2 This dataset is the bearing transmission system fault simulation test bench corresponding to the dataset. The dataset simulates the health state of rolling bearings and the inner ring, outer ring, rolling elements and compound fault states through the SQI test bench. It provides vibration signals collected under steady-state speed conditions of 1200rpm~4800rpm and time-varying speed conditions of 0-2400rpm-0. The data acquisition sampling frequency is set to 25.6kHz.

[0062] Step S1: First, load the vibration signals under the 0-2400rpm-0 condition from the predefined file path (i.e., there are 4 fault state signals and 1 healthy state signal for a 5-classification task; the healthy state is the normal state). Then, load each vibration signal... Z-score normalization is performed to obtain the normalized signal. :

[0063] ;

[0064] in, and These represent the mean and standard deviation of the original vibration signal, respectively.

[0065] Then, a sliding window of length 1024 is used to process the normalized signal. The sample set can be represented as follows: Subsamples are divided without overlap, and each subsample is assigned a corresponding state label. The preset state types and the number of subsamples are shown in Table 1.

[0066] Table 1

[0067]

[0068] Step S2 involves constructing a graph structure for the sample sets under different states, forming a graph structure dataset. The construction of the graph structure for each state specifically involves: starting from the sample set... Each time, 10 subsamples are selected as the node set of a single image. The RadiusGraph method is used to calculate the cosine similarity between nodes. If the cosine similarity between two nodes is greater than a threshold, then... If an edge connects two nodes, then the adjacency matrix of the graph can be obtained. ( The dimension of the adjacency matrix is ​​. ),Right now:

[0069] ;

[0070] In the formula, The element in the p-th row and q-th column of the adjacency matrix is ​​used to determine the node. With nodes To determine whether there is an edge connecting the two sides, a value of 1 indicates that an edge exists, and a value of 0 indicates that no edge exists; p,q=1,2,...,10; The cosine similarity between two nodes is expressed as:

[0071] ;

[0072] in, This is represented as the magnitude of the node feature vector. In this embodiment, the threshold is... Setting it to 0 considers only the positive correlations between nodes when constructing graph-structured data.

[0073] Based on the above, 100 graph structures were constructed for each state, resulting in a total of 500 graph structures, forming a graph structure dataset. Subsequently, the graph structure dataset was divided into a training set and a test set in a 4:1 ratio using stratified sampling to ensure that the proportion of samples in each category was consistent. Figure 3 This is a schematic diagram of the vibration signal sample division and correlation diagram construction process according to an embodiment of the present invention.

[0074] Step S3, Reference Figure 4 For the signals of each node in the graph structure Discrete wavelet transform is used for multi-scale decomposition, first with a low-pass filter. With high-pass filter Convolution is performed, followed by downsampling with a stride of 2 to obtain the low-frequency components. With high frequency components The process can be represented as follows:

[0075] ;

[0076] in, Indicates downsampling, such that The sequence length of the low-frequency and high-frequency components is approximately half of the original length of 1024, meaning that the range of values ​​for n is: Subsequently, based on learnable attention weights... The low-frequency components of each node are enhanced to suppress the interference of noise and local disturbances on the feature representation under varying operating conditions.

[0077] ;

[0078] in, yes The low-frequency component after attention enhancement. In this embodiment, such as Figure 4 As shown, attention weights are estimated using a multilayer perceptron (MLP) and a sigmoid function. , represented as:

[0079] ω p = σ ( MLP ([ LF p , HF p ])) ;

[0080] in, [ LF p , HF p ] express The splicing of low-frequency and high-frequency components along the feature dimension; The sigmoid function can be defined as:

[0081] ;

[0082] Step S4, refer to Figure 5 The low-frequency components of each node after attention enhancement and adjacency matrix The graph structure formed The input consists of stacked graph wavelet activation convolutional layers. Each layer uses a wavelet Kolmogorov-Arnold (WKA) operator (WKA) constructed from wavelet basis functions with learnable scaling factor a and translation factor b to perform nonlinear activation on the features of each node and achieve message passing, thereby extracting node feature representations that are insensitive to changes in operating conditions and have interpretability.

[0083] ;

[0084] in, Indicates the first WKA operator for layer GWAConv; Representative node In the The feature vector of layer GWAConv, its initial value for The feature representation, namely: ; This represents the feature representation of a node after being updated by GWAConv at the k-th layer. Represents a node All adjacent nodes The set, namely: ; Represents a node In the Feature vectors of layer GWAConv; This is a learnable balancing parameter used to adjust the weight of a node's own features during the update process.

[0085] In this embodiment, the constructed interpretable graph wavelet activation network consists of two layers of graph wavelet activation convolutional layers, GWAConv. Each GWAConv layer contains a WKA operator for feature activation and message passing at each input node. The WKA operator is constructed based on the Kolmogorov–Arnold representation theorem, and its mother wavelet function... Using the Mexican hat wavelet, its expression is:

[0086] ;

[0087] Where C is a constant; z represents the feature representation of each input node, i.e., the enhanced low-frequency component. ; t is the independent variable in the mother wavelet function, that is: ; This represents a wavelet basis function with a learnable scaling factor a and a translation factor b; The function composition operation involves cascading multiple wavelet basis functions sequentially and applying them to the features of each input node; that is, the input first passes through the basis functions. The nonlinear mapping, then The output is used as a basis function. The input is repeated, and so on, until the last layer. This achieves multi-layer wavelet nonlinear activation mapping of the features of each input node. In this embodiment, the constructed WKA operator uses only one layer of wavelet basis functions, meaning that the features of each input node undergo only a single layer of wavelet nonlinear mapping, thereby achieving feature activation and multi-scale response. Therefore, the feature update formula for each node in the WKA layer can be simplified to:

[0088] ;

[0089] in, Indicates the first The wavelet basis functions of the WKA operator in layer GWAConv are used for nonlinear activation and message passing of features at each node.

[0090] Therefore, the first layer GWAConv of the constructed interpretable graph wavelet activation network can be expressed as:

[0091] ;

[0092] in, and They represent and its adjacent nodes The initial feature vector; express The output feature vector after the first layer of GWAConv; Let represent the wavelet basis function of the WKA operator in the first layer GWAConv.

[0093] The second layer of the network, GWAConv, can be represented as:

[0094] ;

[0095] in, Represents a node Adjacent nodes The output feature vector after the first layer of GWAConv; Represents a node The output feature vector after the second layer of GWAConv; Let H represent the wavelet basis function of the WKA operator in the second layer GWAConv. Therefore, the node feature representation H extracted by the stacked graph wavelet activation convolutional layers can be defined as:

[0096] .

[0097] Step S5, as follows Figure 6 As shown, the learned node feature representation H is first folded into a graph representation using the Readout operation. The information is then input into the fully connected layer to identify the fault type of the transmission system under variable speed conditions. This process can be represented as follows:

[0098] ;

[0099] ;

[0100] in, Indicates a fully connected layer. It is a non-linear activation function. Used to map the output of the fault classifier to the probability of a state. .

[0101] Figure 7 This is a schematic diagram illustrating the diagnostic results of a fault experiment on a test set for the bearing transmission system according to an embodiment of the present invention. The experimental results compare the proposed Interpretable Graph Wavelet Activation Network (GWAN) with eight common models: Graph Convolutional Network (GCN), Graph Isomorphic Network (GIN), Graph Sampling Aggregation Network (GraphSage), High-Order Graph Neural Network (HoGNN), Convolutional Neural Network (CNN), Multilayer Perceptron (MLP), Kolmogorov-Arnold Network (KAN), and Transformer Network. Table 2 shows the maximum, minimum, and mean values ​​achieved by each model in five experiments.

[0102] from Figure 7 As shown in Table 2, the experimental results demonstrate that the interpretable graph wavelet activation network proposed in this invention achieves optimal fault identification performance in multiple independent experiments. Compared with traditional graph convolutional models (such as GCN, GIN, GraphSAGE, and HoGNN), this invention exhibits superior performance in terms of average accuracy and stability, indicating stronger adaptability in modeling the correlation between vibration signal sub-samples and representing graph node features. Furthermore, compared to models like CNN and MLP that learn independent features from single samples, this invention can achieve information interaction and aggregation between multiple sub-samples through graph structures, thereby obtaining more robust feature representations. On this basis, compared with the Transformer model based on self-attention, this invention maintains high identification accuracy even under non-stationary conditions such as variable speed rotation, further verifying its robustness and practicality under complex working conditions.

[0103] Table 2

[0104]

[0105] like Figure 8As shown in (a), after the original node signals are processed step-by-step in the wavelet attention layer and the GWAConv layer, the fault-related features are gradually highlighted and enhanced. To verify the rationality of the feature aggregation process, the features of all nodes in a randomly selected graph structure are summed, and the cosine similarity is calculated between this summed feature and the node feature representation output by the final wavelet activation convolutional layer in the interpretable graph wavelet activation network. The results show that the summed node features are highly correlated with the node feature representation output by the network, confirming that the network has a clear physical and structural basis when aggregating neighborhood information to construct representations. Secondly, as Figure 8 As shown in (b), the scaling and translation factors of the wavelet basis functions in the network can be adaptively adjusted during training, which can effectively preserve the original waveform of the features. This allows the features after wavelet activation to form a clear correspondence with specific fault modes.

[0106] In summary, the proposed method not only effectively improves the expressive power of fault features, but also provides interpretable evidence for diagnostic results, which helps to realize fault source analysis and reliable decision-making in real industrial scenarios.

[0107] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for diagnosing variable speed faults in a transmission system based on an interpretable graph wavelet activation network, characterized in that, include: Step 1: Collect vibration signals of the transmission system under different states during variable speed operation; normalize each vibration signal, divide the normalized signal into sub-samples of length L, and assign a corresponding state label to each sub-sample of each vibration signal. The resulting sample set is represented as follows: The status includes healthy status and fault status. This represents the i-th subsample of the j-th vibration signal. Then it means The corresponding status label, This represents the number of subsamples of the j-th vibration signal; Step 2: Construct a graph structure for the sample sets under different states; Step 3: Construct an interpretable graph wavelet activation network. This network includes a wavelet attention layer, stacked graph wavelet activation convolutional layers, and a Readout operation. Each node in each graph structure is processed through the wavelet attention layer to obtain enhanced low-frequency components. The enhanced low-frequency components and the graph structure formed by the adjacency matrix are used as input to the graph wavelet activation convolutional layers to obtain the updated feature representation H of the nodes. The Readout operation is then used to fold the updated node feature representation H into a graph representation. ; Step 4, Represent the diagram Input is fed into the fully connected layer to enable state recognition of the transmission system under variable speed conditions.

2. The method for diagnosing variable speed faults in a transmission system based on an interpretable graph wavelet activation network according to claim 1, characterized in that, The step of dividing the normalized signal into subsamples of length L specifically involves using a sliding window of preset length to divide the normalized signal into subsamples without overlap.

3. The method for fault diagnosis of transmission systems based on interpretable graph wavelet activation networks according to claim 1, characterized in that, The construction of the graph structure for the sample set under different states specifically involves: starting from the sample set... Each selection Each subsample serves as a node set for a single image. The RadiusGraph method is used to calculate the cosine similarity between nodes. If the cosine similarity between two nodes is greater than a threshold, it is determined that there is an edge connecting the two nodes, thus obtaining the adjacency matrix of the graph structure. .

4. The method for diagnosing variable speed faults in a transmission system based on an interpretable graph wavelet activation network according to claim 1, characterized in that, The wavelet attention layer is specifically as follows: For each node in each graph structure Discrete wavelet transform is used for multi-scale decomposition to obtain the low-frequency components. With high frequency components ; Based on each node The low-frequency and high-frequency components are based on learnable attention weights. The low-frequency components are enhanced to obtain the enhanced low-frequency components: ； In the formula, It is a node Low-frequency components enhanced with attention.

5. The method for fault diagnosis of variable speed transmission system based on interpretable graph wavelet activation network according to claim 1, characterized in that, The stacked graph wavelet activation convolutional layer comprises two graph wavelet activation convolutional layers, which convert the low-frequency components of each node after attention enhancement. Adjacency matrix of graph structure The graph structure formed As input to the stacked graph wavelet activation convolutional layers, each layer performs feature activation and message passing on each node using the WKA operator constructed from wavelet basis functions with learnable scaling and translation factors, ultimately obtaining the updated feature representation of the nodes. The feature update process for each node in the graph wavelet activation convolutional layer is defined as follows: ； in, Indicates the first WKA operator for layer GWAConv; Representative node In the Feature vectors of layer-map wavelet-activated convolutional layers; For nodes The feature representation updated after the k-th layer of graph wavelet activation convolutional layer; Represents a node All adjacent nodes The set, i.e. ; Represents a node In the Feature vectors of layer-map wavelet-activated convolutional layers; These are learnable equilibrium parameters.

6. The method for fault diagnosis of variable speed transmission system based on interpretable graph wavelet activation network according to claim 1, characterized in that, The diagram representation The input is fed into the fully connected layer to achieve state recognition of the transmission system under variable speed conditions. The expression is: ； in, Indicates a fully connected layer. It is a non-linear activation function. Used to map the output of the fault classifier to the probability of a state. .

7. A transmission system variable speed fault diagnosis system based on an interpretable graph wavelet activation network, characterized in that, The module includes the method described in any one of claims 1-6.

8. A terminal, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to perform the steps of the method according to any one of claims 1-6.

Images