Fault diagnosis method and system based on wavelet graph convolution

Abstract

The present application provides a fault diagnosis method and system based on wavelet graph convolution. The fault diagnosis method based on wavelet graph convolution includes: extracting multi-scale time-frequency features of a signal based on Dobesi wavelets and obtaining wavelet components at multiple scales; performing time-frequency analysis on each wavelet component at each scale and obtaining corresponding wavelet coefficients and time-frequency domain information; constructing a multi-scale graph convolution network combined with a similarity attention mechanism and extracting multi-scale time-frequency graph features; fusing multi-scale time-frequency graph features and obtaining a multi-scale fusion feature; and obtaining a label for a composite fault. This fault diagnosis method based on wavelet graph convolution can solve the technical problem that the fault impacts generated by different fault sources may have complex nonlinear and strong coupling relationships, resulting in low accuracy and speed of fault diagnosis in existing gearbox fault diagnosis methods.

Description

Technical Field

[0001] The present application relates to the technical field of fault diagnosis of rotating machinery parts, and in particular to a fault diagnosis method and system based on wavelet graph convolution. Background Art

[0002] Gearboxes are crucial transmission components in mechanical systems of large equipment like wind turbines, due to their strong load-bearing capacity, compact structure, smooth and accurate power transmission, and high transmission efficiency. They are widely used in large and complex machinery such as wind turbines, helicopters, automobiles, agricultural machinery, and metallurgical machinery. By using gearboxes to change speed and torque, mechanical equipment can achieve functions such as power transmission, speed change, and direction adjustment.

[0003] In complex systems, the probability and severity of failure are higher than in simpler and random systems. Failures can be random, secondary, concurrent, explosive, and hidden, and the occurrence or presence of multiple failures is common. Due to harsh operating environments, heavy loads, high speeds, and long-term continuous operation, typical gearbox components such as gears and rolling bearings are prone to various types of failures. Gear failures typically occur on a single tooth surface or a small number of tooth surfaces and are often compound failures.

[0004] In the study of complex faults, commonly used complex fault methods include statistical methods: based on probability statistics theory, using data analysis and modeling technology to diagnose complex faults. Common methods include Bayesian networks, hidden Markov models, etc. Artificial intelligence methods: based on machine learning and data mining technology, learn and establish models from large amounts of data to achieve the diagnosis of complex faults. Common methods include neural networks, support vector machines, decision trees, etc. Fault tree analysis method: by analyzing the possibility and impact of the occurrence of faults, a fault tree model is established, and qualitative and quantitative analysis of complex faults is performed to find the root cause of the fault. Signal processing method: Use signal processing technology to analyze and extract complex fault signals, such as wavelet transform, time-frequency analysis and other methods, to achieve fault diagnosis and prediction. Physical model method: based on the physical model of the equipment or system, the diagnosis and prediction of complex faults are achieved by analyzing and simulating the model. Common methods include finite element method, multi-body dynamics model, etc.

[0005] When faced with complex gearbox faults, traditional fault diagnosis methods often rely on signal processing, expert experience, and shallow learning to extract features. Feature selection directly impacts diagnostic effectiveness and makes it difficult to fully utilize massive amounts of data. Consequently, gearbox fault diagnosis suffers from issues such as low accuracy and speed, as well as complex nonlinear and strongly coupled relationships between fault impacts from different fault sources, which can severely impact diagnostic accuracy. Summary of the Invention

[0006] The present application provides a fault diagnosis method and system based on wavelet graph convolution to solve the technical problem that the fault impacts generated by different fault sources may have complex nonlinear and strong coupling relationships, resulting in low accuracy and speed of fault diagnosis in existing gearbox fault diagnosis methods.

[0007] The present application provides a fault diagnosis method based on wavelet graph convolution, including:

[0008] Extract the multi-scale time-frequency characteristics of the signal based on Dobesi wavelet and obtain the wavelet components of multiple scales;

[0009] Perform time-frequency analysis on the wavelet components of each scale and obtain the corresponding wavelet coefficients and time-frequency domain information;

[0010] Construct a multi-scale graph convolutional network combined with a similarity attention mechanism and extract multi-scale time-frequency graph features;

[0011] Fusing multi-scale time-frequency graph features and obtaining a multi-scale fusion feature;

[0012] Gets the label of the compound fault.

[0013] Optionally, before the step of extracting multi-scale time-frequency features of the signal based on Dobesi wavelet, the method further includes:

[0014] The original signal is denoised using wavelet.

[0015] Optionally, the step of performing denoising processing on the original signal using wavelet includes:

[0016] Perform wavelet decomposition on the original signal to obtain its wavelet coefficients and approximate coefficients;

[0017] Determine whether the wavelet coefficient is greater than a threshold, if so, keep the wavelet coefficient, if not, set the wavelet coefficient to 0;

[0018] Perform inverse wavelet transform on the filtered wavelet coefficients and approximate coefficients.

[0019] Optionally, the step of extracting multi-scale time-frequency features of the signal based on Dobesi wavelet and obtaining wavelet components of multiple scales includes:

[0020] Based on the Dobesi wavelet decomposition signal x(t), and obtain wavelet components of multiple scales, the decomposition formula is:

[0021]

[0022] in, and represents the low-pass filter and high-pass filter of the db6 wavelet, ψ j,k (t) and φ j,k (t) represents the wavelet basis function, J represents the number of layers of wavelet decomposition, and N represents the length of the filter.

[0023] Optionally, the step of performing time-frequency analysis on the wavelet component of each scale and obtaining corresponding wavelet coefficients and time-frequency domain information includes:

[0024] Perform time-frequency analysis on the wavelet components of each scale and obtain the corresponding wavelet coefficients. The calculation formula of the wavelet coefficients is:

[0025]

[0026] Among them, c i,k represents the wavelet coefficient, j represents the number of layers of wavelet decomposition, k represents the subscript of the wavelet coefficient, ψ j,k (n) represents the discrete form of the wavelet basis function in the time domain;

[0027] The time-frequency domain information of the signal at different scales is obtained based on the wavelet coefficients.

[0028] Optionally, the steps of constructing a multi-scale graph convolutional network combined with a similarity attention mechanism and extracting multi-scale time-frequency graph features include:

[0029] Use the attention mechanism to calculate the attention weight a between nodes i,j , weight coefficient a i,j The expression is:

[0030]

[0031] Among them, a i,j represents the calculated weight between nodes i and j, node k is the neighbor node of node i, N represents the set of neighbor nodes of node i, W represents the weight matrix to be learned, which is used to perform weighted summation of the features of neighbor nodes, and h i , h j , h k Represent the mapping outputs of the i, j, and k layers respectively;

[0032] Get the new i-th layer mapping layer output, the expression is:

[0033] h i =σ∑ j∈N a i,j Wh j

[0034] Among them, h i , h j Represents the mapping output of layer i and layer j respectively, ai,j represents the calculated weight between nodes i and j, and W represents the weight matrix to be learned;

[0035] H parallel convolutional layers are used to simultaneously learn the multi-scale features of the input signal. The output of each convolutional layer can be expressed as:

[0036]

[0037] Among them, i It represents the output of the i-th convolutional layer. represents the convolution operation, and β(·) represents the activation function operation, which performs nonlinear transformation on neurons.

[0038] Optionally, the step of fusing multi-scale time-frequency graph features to obtain a multi-scale fusion feature includes:

[0039] The multi-scale time-frequency graph features are fused to obtain a multi-scale fusion feature, which is expressed as:

[0040]

[0041] Among them, F1, F2, and F3 represent the time-frequency feature output of each individual convolutional layer. Indicates summation, FL Output Represents the final multi-scale feature fusion output.

[0042] Optionally, the step of obtaining a label of a composite fault includes:

[0043] Based on the normalized exponential function, the scale α is obtained i The feature weights of

[0044]

[0045] in, s is the scaling factor of the Softmax function, which is the sum of the exponential functions of all input elements and is used to ensure that the sum of all output elements is 1. Represents the i-th element of the input vector, with e as the base, and normalizes each element;

[0046] The final classification probability is calculated using the feature weights and fusion output values ​​of each scale:

[0047]

[0048] Determine whether the classification probability is less than a preset threshold. If so, output the label of the compound fault.

[0049] Correspondingly, the present application also provides an electronic device, which includes a memory and a processor, the memory being used to store executable program code; the processor being connected to the memory, and running a computer program corresponding to the executable program code by reading the executable program code, so as to execute the steps of any of the aforementioned fault diagnosis methods based on wavelet graph convolution.

[0050] Correspondingly, the present application also provides a fault diagnosis system based on wavelet graph convolution, which includes the aforementioned electronic device.

[0051] This application provides a fault diagnosis method and system based on wavelet graph convolution. This method uses discrete wavelet transform and multi-scale graph convolution combined with a similarity attention mechanism to decouple gearbox composite faults and output multi-labels. After the original data is denoised using wavelets, multi-scale time-frequency information is obtained through the Multibecy wavelet transform. The complex features extracted from the waveform are fed into a multi-scale graph convolution network combined with a similarity attention mechanism for feature fusion. Finally, multiple Softmax probability classifiers are used to complete the multi-label output of the gearbox composite fault. This algorithm has high recognition accuracy and fast processing speed, can effectively diagnose gearbox faults, and complete multi-label output of composite faults.

[0052] The use of discrete wavelet to extract multi-scale time-frequency information and its application in gearbox composite fault diagnosis can produce multi-label output. At the same time, wavelet transform can extract more complete multi-scale time-frequency information.

[0053] The attention mechanism is used in graph convolutional networks to calculate the similarity between nodes, making the graph convolutional neural network more capable of extracting features, having more powerful computing power, better diagnostic performance, and more robustness to signal scale uncertainty.

[0054] The multi-scale graph convolutional network is used to fuse the extracted multi-scale time-frequency information, resulting in a relatively large receptive field and strong information representation capabilities. In order to design convolution kernels of different sizes in the same convolutional layer, the convolution output results of this layer are superimposed to effectively fuse the multi-scale time-frequency information. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0056] Figure 1 This is a flow chart of the fault diagnosis method based on wavelet graph convolution provided by this application;

[0057] Figure 2 This is a schematic diagram of the structure of the graph in step S400 in the fault diagnosis method based on wavelet graph convolution provided by the present application;

[0058] Figure 3 This is a schematic diagram of outputting composite fault labels in the fault diagnosis method based on wavelet graph convolution provided in this application. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application. In addition, it should be understood that the specific implementation methods described herein are only used to illustrate and explain the present application, and are not used to limit the present application. In the present application, unless otherwise specified, the directional words used, such as "up", "down", "left", and "right", generally refer to the up, down, left, and right of the device in actual use or working state, specifically the drawing direction in the accompanying drawings.

[0060] This application provides a fault diagnosis method and system based on wavelet graph convolution, which are described in detail below. It should be noted that the order of description of the following embodiments does not limit the preferred order of the embodiments of this application. In addition, in the following embodiments, the description of each embodiment has its own focus. For parts not detailed in one embodiment, please refer to the relevant description of other embodiments.

[0061] See also Figure 1 The present application provides a fault diagnosis method based on wavelet graph convolution, which is a new composite fault diagnosis algorithm that aims to efficiently extract complex system features, decouple composite faults, and realize multi-label output, thereby improving the accuracy and efficiency of composite fault diagnosis.

[0062] See also Figure 1 ,The fault diagnosis method based on wavelet graph convolution specifically includes the following steps:

[0063] S100, performing denoising processing on the original signal using wavelet;

[0064] The original gearbox acceleration signal contains a lot of noise, which can easily interfere with subsequent feature extraction. Therefore, a denoising method is needed to pre-process the original signal. Using wavelet to denoise the gearbox acceleration signal can reduce the impact of the noise signal on the real data.

[0065] Step S100 specifically includes the following steps:

[0066] S110, performing wavelet decomposition on the original signal to obtain its wavelet coefficients and approximate coefficients;

[0067] Wavelet denoising uses wavelet transform to decompose the signal into multiple frequency bands, which facilitates the processing and analysis of different frequency components, and then denoises the signal according to the characteristics of each frequency band.

[0068] S120, determining whether the wavelet coefficient is greater than a threshold, if so, retaining the wavelet coefficient, if not, setting the wavelet coefficient to 0;

[0069] After the signal is transformed by wavelet, the wavelet coefficients generated by the signal contain important information of the signal. After the signal is decomposed by wavelet, the wavelet coefficients are larger, and the wavelet coefficients of the noise are smaller, and the wavelet coefficients of the noise are smaller than the wavelet coefficients of the signal. By selecting a suitable threshold, the wavelet coefficients greater than the threshold are considered to be generated by the signal and should be retained, while those less than the threshold are considered to be generated by noise and set to zero to achieve the purpose of denoising.

[0070] Threshold processing is performed on the wavelet coefficients, setting the wavelet coefficients smaller than a certain threshold to 0 and retaining the wavelet coefficients larger than a certain threshold. The purpose of threshold processing is to remove the noise components in the signal and retain the main information of the signal.

[0071] S130, performing inverse wavelet transform on the filtered wavelet coefficients and approximate coefficients;

[0072] The processed wavelet coefficients and approximate coefficients are subjected to inverse wavelet transform to obtain the denoised signal.

[0073] S140 , determining whether the denoised signal contains noise components. If so, repeat steps S110 - S130 . If not, output the denoised signal.

[0074] If there are still noise components in the denoised signal, the above steps can be repeated to perform wavelet transform and threshold processing on the processed signal again until the desired denoising effect is achieved.

[0075] Wavelet denoising uses wavelet transform to decompose the signal into multiple frequency bands, and then denoises the signal based on the characteristics of each frequency band. Compared with traditional filtering methods, wavelet denoising can not only remove noise but also retain the important characteristics of the signal.

[0076] S200, extracting multi-scale time-frequency features of the signal based on Dobesi wavelet, and obtaining wavelet components of multiple scales;

[0077] The Dobesi wavelet is a commonly used wavelet basis function with good time-frequency locality and multi-resolution properties. To obtain richer time-frequency domain information, its multi-resolution property can gradually refine the signal's time-frequency characteristics, thereby improving the accuracy of signal analysis and processing. At the same time, to obtain multi-scale features, each scale corresponds to a different time-frequency characteristic, thus enabling multi-scale expression of signal characteristics and more accurately describing the signal's time-frequency characteristics. Furthermore, given the high complexity of composite fault signals, the Dobesi wavelet is well suited for complex signal analysis.

[0078] The denoised signal x(t) is decomposed based on Dobesi wavelet, and wavelet components of multiple scales are obtained. The decomposition formula is:

[0079]

[0080] in, and represents the low-pass filter and high-pass filter of the db6 wavelet, ψ j,k (t) and φ j,k (t) represents the wavelet basis function, J represents the number of layers of wavelet decomposition, and N represents the length of the filter.

[0081] S300, performing time-frequency analysis on the wavelet components of each scale, and obtaining corresponding wavelet coefficients and time-frequency domain information;

[0082] Discrete wavelet is used to extract multi-scale time-frequency information, which is applied to the composite fault diagnosis of gearboxes and multi-label output. At the same time, wavelet transform can extract more complete multi-scale time-frequency information.

[0083] Step S300 specifically includes the following steps:

[0084] S310: Perform time-frequency analysis on the wavelet components of each scale and obtain the corresponding wavelet coefficients. The calculation formula of the wavelet coefficients is:

[0085]

[0086] Among them, c i,k represents the wavelet coefficient, j represents the number of layers of wavelet decomposition, k represents the subscript of the wavelet coefficient, ψ j,k (n) represents the discrete form of the wavelet basis function in the time domain;

[0087] Based on the wavelet coefficients, we can obtain the time-frequency information of the signal at different scales. In the time domain of the wavelet coefficients, each coefficient corresponds to the approximate or detailed information of the original signal at different scales. In the frequency domain of the wavelet coefficients, each coefficient corresponds to the signal component of different frequencies. Therefore, we can extract the multi-scale time-frequency domain information we need from the wavelet coefficients.

[0088] S320, obtaining time-frequency domain information of the signal at different scales according to the wavelet coefficients;

[0089] The Dobesi wavelet decomposes a signal into wavelet components at multiple scales. Each scale corresponds to a signal component of a different frequency. The time and frequency domain information of the signal component is contained in the wavelet coefficients. By analyzing the wavelet coefficients, we can obtain the multi-scale time-frequency information of the signal, thereby better describing the characteristics of the signal.

[0090] S400: Construct a multi-scale graph convolutional network combined with a similarity attention mechanism and extract multi-scale time-frequency graph features;

[0091] Graph convolutional neural network is a convolutional neural network that can directly act on the graph and utilize its structural information. For each node, we need to consider all its neighbors and the feature information contained in it. Graph convolution can aggregate node features near a node and learn node features through weighted aggregation to perform a series of prediction tasks. The graph structure is represented by G = (V, E), where V represents the geometry of the node, E represents the geometry of the edge, n represents the number of nodes, and m represents the number of edges. Figure 2 As shown, there are n nodes, each with its own feature vector. We assume that the features of these nodes form an n×d-dimensional feature matrix X, and the relationships between the nodes also form an n×n-dimensional adjacency matrix A. X and A are the inputs of our model.

[0092] right Figure 2 The weighted sum of each node's neighboring nodes is taken, and the features of the nodes in the next layer are obtained by multiplying them with the parameter matrix. The features of each node are iteratively calculated in matrix form, and then convolution operations are performed through layer propagation, and finally the node features are updated. However, since only average weighting is used, the similarity between nodes is not taken into account. Therefore, a weight is introduced to consider the impact of node similarity on the system, and the weight between two nodes is calculated using the attention mechanism.

[0093] Step S400 specifically includes the following steps:

[0094] S410, using the attention mechanism to calculate the attention weight a between nodes i,j , weight coefficient a i,j The expression is:

[0095]

[0096] Among them, a i,j represents the calculated weight between nodes i and j, node k is the neighbor node of node i, N represents the set of neighbor nodes of node i, W represents the weight matrix to be learned, which is used to perform weighted summation of the features of neighbor nodes, and h i, h j , h k Represent the mapping outputs of the i, j, and k layers respectively;

[0097] S420, obtain the new i-th mapping layer output, expressed as:

[0098] h i =σ∑ j∈N a i,j Wh j

[0099] Among them, h i , h j Represents the mapping output of layer i and layer j respectively, a i,j represents the calculated weight between nodes i and j, and W represents the weight matrix to be learned;

[0100] S430, using h parallel convolutional layers to simultaneously learn multi-scale features of the input signal, the output of each convolutional layer can be expressed as:

[0101]

[0102] Among them, i It represents the output of the i-th convolutional layer. represents the convolution operation, β(·) represents the activation function operation, which performs nonlinear transformation on neurons;

[0103] Because multi-scale features need to be integrated, multi-scale convolution kernels are the most effective method for multi-scale feature fusion. The advantages of multi-scale feature fusion are a relatively large receptive field and strong information representation capabilities. To design convolution kernels of different sizes within the same convolutional layer, the convolution outputs of that layer are superimposed. Furthermore, through transformations, convolutions and pooling of varying sizes with varying receptive fields can be obtained.

[0104] In addition, the specific operation of the multi-scale convolution kernel is defined as follows:

[0105]

[0106] Among them, W (k-1,k,k+1) Represents three convolution kernels of different scales, X is the set of x, b (k-1,k,k+1) Denotes the offset of the three convolutions. And after each convolution layer, batch normalization and activation function are used.

[0107] S500, fusing multi-scale time-frequency graph features to obtain a multi-scale fusion feature;

[0108] The generated multi-scale graph convolution features are fused, and the final multi-scale fusion feature is expressed as:

[0109]

[0110] Among them, F1, F2, and F3 represent the time-frequency feature output of each individual convolutional layer. Indicates summation, FL Output Represents the final multi-scale feature fusion output.

[0111] The extracted multi-scale time-frequency graph features are put into the multi-scale graph convolutional network as input, so that the extracted features have a relatively large receptive field and strong information representation ability. At the same time, combined with the similarity attention mechanism, the similarity of node information in the graph structure is taken into account to achieve the fusion of multi-scale features.

[0112] The attention mechanism is used in graph convolutional networks to calculate the similarity between nodes, making the graph convolutional neural network more capable of extracting features, having more powerful computing power, better diagnostic performance, and more robustness to signal scale uncertainty.

[0113] See also Figure 3 ,S600, obtain the label of the composite fault;

[0114] The multi-label output of gearbox composite faults is completed through multiple Softmax probability classifiers, which specifically includes the following steps:

[0115] S610, based on the normalized exponential function, obtain the scale α i The feature weights of

[0116]

[0117] in, s is the scaling factor of the Softmax function, which is the sum of the exponential functions of all input elements and is used to ensure that the sum of all output elements is 1. Represents the i-th element of the input vector, with e as the base, and normalizes each element;

[0118] The Softmax function can map an input real number to the range of 0-1, that is, convert our score value into a corresponding probability value. Multiple classifiers can be used simultaneously to complete the multi-label classification of gearbox compound faults.

[0119] S620: Calculate the final classification probability using the feature weights and fusion output values ​​of each scale:

[0120]

[0121] S630: Determine whether the classification probability is less than a preset threshold; if so, output a label of the composite fault;

[0122] The greater the probability of correct classification, the smaller the loss function value. The smaller the probability of correct classification, the larger the loss function value should be. When the probability of correct classification is 1, that is, when the classification result is completely correct, our loss function value should be 0.

[0123] In this embodiment, a threshold value θ=0.5 is preset as a criterion for determining the positive and negative output categories, thereby completing the multi-label output of the gearbox composite fault.

[0124] This embodiment provides a method for decoupling and multi-label output of gearbox composite faults based on discrete wavelet transform and multi-scale graph convolution with a similarity attention mechanism. After raw data is denoised using wavelet transforms, multi-scale time-frequency information is obtained through a Multibecy wavelet transform. Complex features extracted from the waveforms are then fed into a multi-scale graph convolutional network with a similarity attention mechanism for feature fusion. Finally, multiple Softmax probabilistic classifiers are used to achieve multi-label output for composite gearbox faults. This algorithm offers high recognition accuracy and fast processing speed, enabling effective gearbox fault diagnosis and multi-label output for composite faults.

[0125] Compared with traditional fault diagnosis methods, which often extract features through signal processing, expert experience, shallow learning, etc., the selection of features directly affects the fault diagnosis effect, and it is difficult to fully utilize massive data. The fault diagnosis of gearboxes has problems such as low accuracy and speed of fault diagnosis, and the complex nonlinear and strong coupling relationship between the fault impacts generated by different fault sources may seriously affect the accuracy of diagnosis. This embodiment designs a multi-scale graph convolutional network to fuse the extracted multi-scale time-frequency information. The advantage of multi-scale feature fusion is that it has a relatively large receptive field and strong information representation ability. In order to design convolution kernels of different sizes in the same convolution layer, the results of the convolution output of this layer are superimposed, which can effectively fuse multi-scale time-frequency information.

[0126] The present application also provides a fault diagnosis system based on wavelet graph convolution, which includes an electronic device, which includes a memory and a processor, the memory being used to store executable program code; the processor being connected to the memory, and running a computer program corresponding to the executable program code by reading the executable program code to execute the steps of any of the aforementioned fault diagnosis methods based on wavelet graph convolution.

[0127] The above is a detailed introduction to the fault diagnosis method and system based on wavelet graph convolution provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, based on the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A fault diagnosis method based on wavelet graph convolution, characterized in that: include: Extract the multi-scale time-frequency characteristics of the signal based on Dobesi wavelet and obtain the wavelet components of multiple scales; Perform time-frequency analysis on the wavelet components of each scale and obtain the corresponding wavelet coefficients and time-frequency domain information; Construct a multi-scale graph convolutional network combined with a similarity attention mechanism and extract multi-scale time-frequency graph features; Fusing multi-scale time-frequency graph features and obtaining a multi-scale fusion feature; Get the label of the compound fault, The steps of constructing a multi-scale graph convolutional network combined with a similarity attention mechanism and extracting multi-scale time-frequency graph features include: Use the attention mechanism to calculate the attention weight a between nodes i,j , weight coefficient a i,j The expression is: Among them, a i,j represents the calculated weight between nodes i and j, node k is the neighbor node of node i, N represents the set of neighbor nodes of node i, W represents the weight matrix to be learned, which is used to perform weighted summation of the features of neighbor nodes, and h i , h j , h k Represent the mapping outputs of the i, j, and k layers respectively; Get the new i-th layer mapping layer output, the expression is: h i =σ∑ j∈N a i,j Wh j Among them, h i , h j Represents the mapping output of layer i and layer j respectively, a i,j represents the calculated weight between nodes i and j, and W represents the weight matrix to be learned; h parallel convolutional layers are used to simultaneously learn the multi-scale features of the input signal. The output of each convolutional layer is expressed as: Among them, i It represents the output of the i-th convolutional layer. represents the convolution operation, and β(·) represents the activation function operation, which performs nonlinear transformation on neurons.

2. The fault diagnosis method based on wavelet graph convolution according to claim 1, characterized in that: Before the step of extracting the multi-scale time-frequency features of the signal based on the Dobesi wavelet, the method further includes: The original signal is denoised using wavelet.

3. The fault diagnosis method based on wavelet graph convolution according to claim 2 is characterized in that: The step of performing denoising processing on the original signal using wavelet comprises: Perform wavelet decomposition on the original signal to obtain its wavelet coefficients and approximate coefficients; Determine whether the wavelet coefficient is greater than a threshold, if so, keep the wavelet coefficient, if not, set the wavelet coefficient to 0; Perform inverse wavelet transform on the filtered wavelet coefficients and approximate coefficients.

4. The fault diagnosis method based on wavelet graph convolution according to claim 1, characterized in that: The step of extracting multi-scale time-frequency features of a signal based on Dobesi wavelet and obtaining wavelet components of multiple scales includes: Based on the Dobesi wavelet decomposition signal x(t), and obtain wavelet components of multiple scales, the decomposition formula is: in, and represents the low-pass filter and high-pass filter of the db6 wavelet, ψ j,k (t) and φ j,k (t) represents the wavelet basis function, J represents the number of layers of wavelet decomposition, and N represents the length of the filter.

5. The fault diagnosis method based on wavelet graph convolution according to claim 1, characterized in that: The steps of performing time-frequency analysis on the wavelet components of each scale and obtaining corresponding wavelet coefficients and time-frequency domain information include: Perform time-frequency analysis on the wavelet components of each scale and obtain the corresponding wavelet coefficients. The calculation formula of the wavelet coefficients is: Among them, c i,k represents the wavelet coefficient, j represents the number of layers of wavelet decomposition, k represents the subscript of the wavelet coefficient, ψ j,k (n) represents the discrete form of the wavelet basis function in the time domain; The time-frequency domain information of the signal at different scales is obtained based on the wavelet coefficients.

6. The fault diagnosis method based on wavelet graph convolution according to claim 1, characterized in that: The step of fusing multi-scale time-frequency graph features and obtaining a multi-scale fusion feature includes: The multi-scale time-frequency graph features are fused to obtain a multi-scale fusion feature, which is expressed as: Among them, F1, F2, and F3 represent the time-frequency feature output of each individual convolutional layer. Indicates summation, FL Output Represents the final multi-scale feature fusion output.

7. The fault diagnosis method based on wavelet graph convolution according to claim 1, characterized in that: The step of obtaining a label of a composite fault includes: Based on the normalized exponential function, the scale α is obtained i The feature weights of in, s is the scaling factor of the Softmax function, which is the sum of the exponential functions of all input elements and is used to ensure that the sum of all output elements is 1. Represents the i-th element of the input vector, with e as the base, and normalizes each element; The final classification probability is calculated using the feature weights and fusion output values ​​of each scale: Determine whether the classification probability is less than a preset threshold. If so, output the label of the compound fault.

8. An electronic device, characterized in that: include: a memory for storing executable program code; as well as A processor is connected to the memory and runs a computer program corresponding to the executable program code by reading the executable program code to execute the steps of the fault diagnosis method based on wavelet graph convolution according to any one of claims 1 to 7.

9. A fault diagnosis system based on wavelet graph convolution, characterized in that: The electronic device comprising the electronic device according to claim 8.

Images