Transmission system fault diagnosis method, system, medium and equipment

Abstract

The invention discloses a transmission system fault diagnosis method and system, a medium and equipment, and the method comprises the steps: collecting a multi-sensor signal of a transmission system during the gear fault of a time-varying rotating speed and a time-varying load; constructing a multi-sensor network time-varying graph data set, and dividing a training set and a test set; inputting the training set data into a spectrogram wavelet packet convolutional neural network for training to realize fault classification; and testing the trained spectrogram wavelet packet convolutional neural network by using test set data, and storing the spectrogram wavelet packet convolutional neural network with the best test performance for transmission system fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of transmission system fault diagnosis, and in particular to a transmission system fault diagnosis method, system, medium and equipment. Background Art

[0002] The transmission system is a crucial execution system in completing the intelligent automation of modern high-end equipment, and its health status has a huge impact on the operating safety and equipment reliability of major equipment. The gearbox, as the most common component in the rotating mechanical system, is subjected to harsh conditions such as time-varying working conditions, heavy loads, high temperatures, dust, etc., and is very prone to failure. Therefore, it is of great significance to detect faults in the gearbox of the transmission system. Existing fault diagnosis methods usually use traditional convolutional neural networks to learn and train single or multiple spliced ​​sensor signals, and the collected fault signals often work under stable working conditions. However, it is difficult for traditional convolutional neural networks to explore the relationship between multi-sensor signals to enhance the diagnostic effect. In addition, once the working conditions change, the performance of traditional networks will also decline.

[0003] The above information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the invention

[0004] The present invention provides a transmission system fault diagnosis method, system, medium and equipment, which can effectively mine spatial information such as redundant information and difference information in multi-sensor network signals to enhance the diagnosis effect; can well perform refined spectral feature transformation on non-steady-state time-varying graph signals, and by combining the two, can realize refined adaptive spectral filtering and frequency band selection for non-steady-state variable operating condition data, thereby enhancing the model's diagnostic ability for variable operating condition data.

[0005] A transmission system fault diagnosis method includes:

[0006] Step S100: collecting multi-sensor signals of the transmission system when a gear fails with time-varying speed and time-varying load;

[0007] Step S200: construct a multi-sensor network time-varying graph dataset and divide it into a training set and a test set;

[0008] Step S300: input the training set data into the spectral wavelet packet convolutional neural network for training to achieve fault classification;

[0009] Step S400: Use the test set data to test the trained spectral wavelet packet convolutional neural network, and save the spectral wavelet packet convolutional neural network with the best test performance for transmission system fault diagnosis.

[0010] In the transmission system fault diagnosis method, in step S100, the multi-sensor signals include vibration, rotation speed, torque and current signals.

[0011] In the transmission system fault diagnosis method, in step S200, constructing a multi-sensor network time-varying graph data set includes:

[0012] S201: For a multivariate time series consisting of τ sensors X = [X 1 , X 2 , …, X τ ], for a single sensor X i The univariate sequence is divided into sub-samples of length l and a corresponding label is assigned to each sub-sample. There is no overlap between samples. The obtained sub-sample set is expressed as:

[0013] ,

[0014] In the formula, ∏ is the obtained subsample set, x represents the subsample, y represents the label, and n represents the number of subsamples;

[0015] S202: According to the obtained sub-sample set Ώ, each sub-sample is regarded as a node of the multi-sensor graph network to construct a graph with τ nodes, and the sensor neighbors are found by calculating the cosine distance. The adjacency matrix is ​​obtained by the RadiusGraph method, including:

[0016] ,

[0017] Where -radius(·) is the calculation node x i and node x j Cosine similarity of , A(i,j) is the value of the i-th row and j-th column in the adjacency matrix of the multi-sensor graph, To determine whether there is an edge between nodes, the threshold is set to 0, and the node feature matrix V is obtained by dividing the sub-sample set Obtained by node correspondence.

[0018] In the transmission system fault diagnosis method, in step S300, the spectral wavelet packet convolutional neural network is obtained by superimposing two spectral wavelet packet convolutional layers to extract multi-scale features of the graph spectrum and two fully connected layers to perform fault classification, which includes the following steps:

[0019] S301: The two spectral wavelet convolutional layers of the spectral wavelet packet convolutional neural network are defined as follows:

[0020] ,

[0021] Among them, V and H respectively represent the input graph node feature vector and the node representation after learning. and are respectively the learnable filter coefficient matrices in two spectral graph wavelet convolutional layers, Relu( ) is a non-linear activation function. is the spectral graph wavelet packet transform operator, which is obtained through the graph adjacency matrix A and degree matrix D, and T represents the transpose.

[0022] S302: Send the output of the second spectral graph wavelet packet convolutional layer into the graph readout layer, and the final graph representation is:

[0023] ,

[0024] where N is the number of graph nodes. is the final graph representation.

[0025] S303: Input the obtained fused graph representation features into the fully connected layer FCLayers for classification, and the predicted probability of the fault mode is:

[0026] ,

[0027] where softmax( ) normalizes the output of the fully connected layer to the probability distribution of the predicted output classes, and FCLayers represents a single-layer fully connected layer.

[0028] S304: According to the obtained predicted probability of the fault mode, use the cross-entropy loss function to train the model, which is expressed as:

[0029] ,

[0030] where M is the number of graphs. is the number of fault categories. is the sign function, which is 0 or 1. If the true class of the i-th sample is equal to take 1, otherwise take 0. is the predicted probability that the i-th sample belongs to the class .

[0031] In the described transmission system fault diagnosis method, in step S301, the spectral graph wavelet packet transform operator is obtained through the graph adjacency matrix A and degree matrix D, and the steps are as follows:

[0032] S3011: First, obtain the Laplacian matrix L according to the graph adjacency matrix A and degree matrix D:

[0033] ,

[0034] S3012: Any filter constructed by the spectrogram wavelet packet transform is the initial scaling function filter and the initial wavelet function filter is the dilation product of, and its dilation is expressed as , where is the dilation scale factor. According to the definition of the spatial decomposition of the spectrogram wavelet packet, the first three layers of spectrogram wavelet packet operators are obtained:

[0035]

[0036] The number of layers of the spectrogram wavelet packet continues to decompose as the dilation factor dilates. Only the first three layers are shown here.

[0037] S3013: Using the scale kernel function approximated by the Chebyshev polynomial and the wavelet kernel function :

[0038]

[0039] Among them, is the eigenvalue obtained by the L decomposition of the Laplacian matrix, , is the 0th coefficient of expressed by the Chebyshev polynomial and , , is the kth coefficient of expressed by the Chebyshev polynomial and , is the shifted Chebyshev polynomial with the domain , and its iterative form is:

[0040]

[0041] Among them, is the largest eigenvalue, obtained by the Arnoldi algorithm. The kth coefficient of the shifted Chebyshev polynomial:

[0042]

[0043] Among them, is the independent variable of the function, is the pi.

[0044] S3014: After obtaining the Chebyshev polynomial approximation coefficients of the scale kernel function and wavelet kernel function at each expansion coefficient for the Chebyshev polynomial approximation, obtain the Chebyshev polynomial coefficients after multiplying two Chebyshev polynomials according to the following formula. Two polynomials in Chebyshev form:

[0045] ,

[0046] where, , the product of polynomials has the following Chebyshev form:

[0047]

[0048] The Chebyshev coefficients of the new product polynomial are obtained through the following formula:

[0049] ,

[0050] Based on this, obtain the spectral graph wavelet packet transform operator for the first layer and subsequent layers , and apply it to the spectral graph wavelet packet convolutional neural network.

[0051] In the described transmission system fault diagnosis method, in step S303, the fully connected layer FCLayers is represented as follows:

[0052] ,

[0053] where, is the weight matrix to be learned, is the bias vector.

[0054] In the described transmission system fault diagnosis method, in step S400, after each pair of spectral graph wavelet packet convolutional neural networks uses all the training set data to train and update the weights, the spectral graph wavelet packet convolutional neural network diagnosis accuracy is tested using the test set data, and the test accuracy rate is obtained by dividing the number of correctly classified samples in the test set by the total number of samples. The model undergoes multiple rounds of training and learning, and the spectral graph wavelet packet convolutional neural network with the highest test accuracy rate is saved as the spectral graph wavelet packet convolutional neural network for variable working condition fault diagnosis.

[0055] A transmission system fault diagnosis system includes,

[0056] An acquisition unit, which is used to acquire multi-sensor signals when the transmission system has gear faults under time-varying rotational speed and time-varying load;

[0057] A construction unit, which is used to construct a time-varying graph data set of a multi-sensor network and divide it into a training set and a test set;

[0058] A training unit, which is used to input training set data into a spectrogram wavelet packet convolutional neural network for training to achieve fault classification;

[0059] A testing unit, which is used to test the trained spectrogram wavelet packet convolutional neural network using test set data, and save the spectrogram wavelet packet convolutional neural network with the best test performance for transmission system fault diagnosis.

[0060] A computer storage medium, the storage medium includes computer instructions, when it runs on a computer, it causes the computer to execute the described method.

[0061] An electronic device, the electronic device includes:

[0062] A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein,

[0063] When the processor executes the program, it implements the described method.

[0064] Compared with the prior art, the graph wavelet method, as an effective method for transforming and extracting features specifically for non-stationary time-varying graph signals in the field of graph signal analysis, can be introduced into the neural network to solve the above existing problems at the same time. On the one hand, the graph neural network can effectively mine spatial information such as redundant information and differential information in multi-sensor network signals to enhance the diagnostic effect; on the other hand, the spectrogram wavelet packet can well perform refined spectral feature transformation on non-stationary time-varying graph signals. By combining the two, refined adaptive spectral filtering and frequency band selection for non-stationary variable working condition data can be realized, and the diagnostic ability of the model for variable working condition data can be enhanced. Description of the Drawings

[0065] By reading the following detailed description of the preferred specific embodiments, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The drawings in the specification are only for the purpose of showing the preferred embodiments, and are not considered to be a limitation of the present invention. Obviously, the following described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Moreover, throughout the drawings, the same reference numerals are used to represent the same components.

[0066] In the drawings:

[0067] Figure 1 is a flowchart of a transmission system fault diagnosis method based on a spectrogram wavelet packet neural network according to an embodiment of the present disclosure;

[0068] Figure 2 is a schematic diagram of the gearbox test bench of the transmission system and the gear fault state provided by an embodiment of the present disclosure;

[0069] Figure 3 is a schematic diagram comparing the model diagnosis accuracy of the model provided by an embodiment of the present disclosure under different signal-to-noise ratio conditions.

[0070] The present invention will be further explained below in conjunction with the accompanying drawings and embodiments. Specific Embodiments

[0071] The specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although specific embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present invention can be more thoroughly understood and the scope of the present invention can be fully conveyed to those skilled in the art.

[0072] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different terms to refer to the same component. The specification and claims do not use the difference in terms as a way to distinguish components, but use the difference in the functions of components as the criterion for distinction. As mentioned throughout the specification and claims, "comprising" or "including" is an open-ended term and should be interpreted as "including but not limited to". The subsequent description of the specification is for the purpose of implementing the preferred embodiments of the present invention, but the description is for the general purpose of the specification and is not intended to limit the scope of the present invention. The protection scope of the present invention shall be defined by the appended claims.

[0073] For the convenience of understanding the embodiments of the present invention, the following will further explain with specific embodiments as examples in conjunction with the accompanying drawings, and each accompanying drawing does not constitute a limitation to the embodiments of the present invention.

[0074] As Figures 1 to 3 shown, the transmission system fault diagnosis method includes the following steps:

[0075] Step S100: Collect multi-sensor signals when the gears of the transmission system are faulty under time-varying rotational speed and time-varying load;

[0076] Step S200: Construct a time-varying graph data set of the multi-sensor network and divide it into a training set and a test set;

[0077] Step S300: Input the training set data into the spectral graph wavelet packet convolutional neural network for training to achieve fault classification;

[0078] Step S400: Use the test set data to test the trained spectrogram wavelet packet convolutional neural network, and save the spectrogram wavelet packet convolutional neural network with the best test performance for transmission system fault diagnosis.

[0079] In the preferred embodiment of the described transmission system fault diagnosis method, in step S100, the multi-sensor signals include vibration, speed, torque, and current signals.

[0080] In the preferred embodiment of the described transmission system fault diagnosis method, in step S200, constructing the multi-sensor network time-varying graph data set includes

[0081] S201: For the multivariate time series X = [X 1 , X 2 , …, X τ composed of τ sensors, divide the univariate sequence of a single sensor X i , divide the original data with length l into sub-samples with length d, and assign corresponding labels to each sub-sample. There is no overlap between samples. The obtained sub-sample set is expressed as:

[0082] ,

[0083] wherein, ∏ is the obtained sub-sample set, x represents the sub-sample, y represents the label, and n represents the number of sub-samples;

[0084] S202: According to the obtained sub-sample set ∏, regard each sub-sample as a node of the multi-sensor graph network to construct a graph with τ nodes, and find sensor neighbors by calculating the cosine distance. The adjacency matrix obtained by the RadiusGraph method includes

[0085] ,

[0086] wherein, -radius(·) calculates the cosine similarity i between node x j and node x , A(i, j) is the value of the i-th row and j-th column in the adjacency matrix of the multi-sensor graph, is the threshold for determining whether there is an edge between nodes, and the threshold is set to 0. The node feature matrix V is obtained according to the node correspondence by dividing the sub-sample set .

[0087] In the preferred embodiment of the described transmission system fault diagnosis method, in step S300, the spectrogram wavelet packet convolutional neural network is obtained by stacking two layers of spectrogram wavelet packet convolutional layers for multi-scale feature extraction of the graph spectrum and two fully connected layers for fault classification, and it includes the following steps:

[0088] S301: The two spectrogram wavelet convolutional layers of the spectrogram wavelet packet convolutional neural network are defined as follows:

[0089] ,

[0090] where V and H represent the input graph node feature vector and the learned node representation respectively, and are the learnable filter coefficient matrices in the two spectrogram wavelet convolutional layers respectively, Relu( ) is the non-linear activation function, is the spectrogram wavelet packet transform operator, obtained through the graph adjacency matrix A and the degree matrix D, and T represents the transpose.

[0091] S302: Send the output of the second spectrogram wavelet packet convolutional layer into the graph readout layer, and the final graph representation is:

[0092] ,

[0093] where N is the number of graph nodes, is the final graph representation,

[0094] S303: Input the obtained fused graph representation features into the fully connected layer FCLayers for classification, and the predicted probability of the fault mode is:

[0095] ,

[0096] where softmax( ) normalizes the output of the fully connected layer into the probability distribution of the predicted output classes, and FCLayers represents a single-layer fully connected layer,

[0097] S304: According to the obtained predicted probability of the fault mode, use the cross-entropy loss function to train the model, which is expressed as:

[0098] ,

[0099] where M is the number of graphs, is the number of fault categories, is the sign function, which is 0 or 1. If the true class of the i-th sample is equal to take 1, otherwise take 0, is the predicted probability that the i-th sample belongs to the class .

[0100] In the preferred embodiment of the described transmission system fault diagnosis method, in step S301, the spectrogram wavelet packet transform operator is obtained through the graph adjacency matrix A and the degree matrix D, and the steps are as follows:

[0101] S3011: First, obtain the Laplacian matrix L based on the adjacency matrix A and degree matrix D of the graph:

[0102] ,

[0103] S3012: Any filter constructed by the spectral graph wavelet packet transform is the dilation product of the initial scaling function filter and the initial wavelet function filter , and its dilation is expressed as , where is the dilation scale factor. According to the spatial decomposition definition of the spectral graph wavelet packet, the first three layers of spectral graph wavelet packet operators are obtained:

[0104]

[0105] The number of layers of the spectral graph wavelet packet continues to decompose with the dilation of the dilation factor. Only the first three layers are shown here.

[0106] S3013: Use the scale kernel function approximated by the Chebyshev polynomial and the wavelet kernel function :

[0107]

[0108] Among them, is the eigenvalue obtained by the eigen-decomposition of the Laplacian matrix L, , is the 0th coefficient of expressed by the Chebyshev polynomial and , , is the kth coefficient of expressed by the Chebyshev polynomial and , is the shifted Chebyshev polynomial with the domain , and its iterative form is:

[0109]

[0110] Among them, is the maximum eigenvalue obtained by the Arnoldi algorithm, and the kth coefficient of the shifted Chebyshev polynomial:

[0111]

[0112] Among them, is the independent variable of the function, is the pi.

[0113] S3014: After obtaining the Chebyshev polynomial approximation coefficients of the scaling kernel function and wavelet kernel function at each dilation coefficient in the Chebyshev polynomial approximation, the Chebyshev polynomial coefficients after multiplying two Chebyshev polynomials are obtained according to the following formula. The two polynomials in Chebyshev form are:

[0114] ,

[0115] where, , the product of polynomials has the following Chebyshev form:

[0116]

[0117] The Chebyshev coefficients of the new product polynomial are obtained by the following formula:

[0118] ,

[0119] Based on this, the spectral graph wavelet packet transform operators of the first layer and subsequent layers are obtained and applied to the spectral graph wavelet packet convolutional neural network.

[0120] In the preferred embodiment of the described transmission system fault diagnosis method, in step S303, the fully connected layer FCLayers is represented as follows:

[0121] ,

[0122] where, is the weight matrix to be learned, is the bias vector.

[0123] In the preferred embodiment of the described transmission system fault diagnosis method, in step S400, after each pair of spectral graph wavelet packet convolutional neural networks uses all the training set data to train and update the weights, the test set data is used to test the diagnostic accuracy of the spectral graph wavelet packet convolutional neural network, and the test accuracy rate is obtained by dividing the number of samples correctly classified on the test set by the total number of samples. The model undergoes multiple rounds of training and learning, and the spectral graph wavelet packet convolutional neural network with the highest test accuracy rate is saved as the spectral graph wavelet packet convolutional neural network for variable working condition fault diagnosis.

[0124] A transmission system fault diagnosis system includes,

[0125] An acquisition unit, which is used to acquire multi-sensor signals of the transmission system during gear faults with time-varying rotational speed and time-varying load;

[0126] A building unit, which is used to build a time-varying graph dataset of a multi-sensor network and divide the training set and the test set;

[0127] A training unit, which is used to input the training set data into a spectral graph wavelet packet convolutional neural network for training to achieve fault classification;

[0128] A testing unit, which is used to test the trained spectral graph wavelet packet convolutional neural network using the test set data, and save the spectral graph wavelet packet convolutional neural network with the best test performance for transmission system fault diagnosis.

[0129] A computer storage medium, the storage medium includes computer instructions, when it runs on a computer, it causes the computer to execute the described method.

[0130] An electronic device, the electronic device includes:

[0131] A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein,

[0132] When the processor executes the program, it implements the described method.

[0133] In one embodiment, as Figure 1 shown, a transmission system fault diagnosis method based on a spectral graph wavelet packet neural network includes the following steps:

[0134] Step S100: Collect multi-sensor signals of the transmission system when there are gear faults under time-varying speed and time-varying load;

[0135] In this embodiment, 4 vibration sensors, 2 current sensors, and 2 rotational speed signals and 2 torque signals collected by 2 encoders are used to simulate the operating states of the transmission system gearbox under normal conditions, planet gear end breakage faults, planet gear spalling faults, and compound faults of planet gear cracks and sun gear cracks. As Figure 2 shown, the data sampling frequency is 10Khz.

[0136] Step S200: Build a time-varying graph dataset of the multi-sensor network and divide the training set and the test set;

[0137] In this embodiment, the signals of 10 of the sensors are used to build multi-sensor graph signals. Each 1024 points of each sensor are intercepted as a node signal, and each graph has 10 nodes for 10 sensors. The node adjacency matrix is obtained through the RadiusGraph method to build a multi-sensor graph signal dataset. Among them, there are 1000 samples for each of the 4 states, with a total of 4000 samples. The samples are divided according to the ratio of training set:test set = 1:1. Therefore, the number of training set samples and the number of test set samples are both 2000.

[0138] Step S300: Input the training set data into the spectrogram wavelet packet convolutional neural network for training to achieve fault classification;

[0139] In this step, the spectrogram wavelet packet convolutional neural network consists of two layers of spectrogram wavelet packet convolutional layers, a graph readout layer, and two layers of fully connected layers. The kernel function uses the Meyer kernel function. In addition, to illustrate the effectiveness of the method, the most classical graph convolutional network model GCN in the graph neural network and the most popular model graph transformer model GraphTransformer in recent years are used as comparison models. The readout layer is the same as the fully connected layer, and only the spectrogram wavelet packet convolutional layer is replaced by the GCN convolutional layer and the Graph Transformer convolutional layer.

[0140] Step S400: Use the test set data to test the trained model and save the model with the best test performance.

[0141] In another embodiment, in step S200, the method for constructing the multi-sensor network time-varying graph data includes the following steps:

[0142] S201: For the multi-variable time series X = [X 1 , X 2 , …, X τ composed of τ sensors, first divide the single-variable sequence of a single sensor X i . Divide the original data with length l into sub-samples with length d, and assign corresponding labels to each sub-sample. There is no overlap between the samples. The obtained sub-sample set can be expressed as:

[0143] ,

[0144] where ∏ is the obtained sub-sample set, x represents the sub-sample, y represents the label, and n represents the number of sub-samples;

[0145] S202: According to the obtained sub-sample set ∏, regard each sub-sample as a node of the multi-sensor graph network to construct a graph with τ nodes, and find the sensor neighbors by calculating the cosine distance, hereinafter denoted as the RadiusGraph method. The method for obtaining the adjacency matrix by the RadiusGraph method can be expressed as follows:

[0146]

[0147] where -radius(·) calculates the cosine similarity i between node x j and node x , A(i, j) is the value of the i-th row and j-th column in the adjacency matrix of the graph, To determine the threshold for whether there is an edge between nodes, in this method, the threshold is set to 0. The node feature matrix V is obtained by partitioning the sub-sample set according to the node correspondence.

[0148] In another embodiment, in step S300, the spectral graph wavelet packet convolutional neural network (SGWPNet) is obtained by stacking two spectral graph wavelet packet convolutional layers (SGWPConv) for multi-scale feature extraction of the refined graph spectrum and two fully connected layers for fault classification, and it includes the following steps:

[0149] S301: The two spectral graph wavelet convolutional layers of the spectral graph wavelet packet convolutional neural network are defined as follows:

[0150]

[0151] where V and H respectively represent the input graph node feature vector and the learned node representation, and are respectively the learnable filter coefficient matrices in the two spectral graph wavelet convolutional layers, and Relu( ) is the non-linear activation function. is the spectral graph wavelet packet transform operator, which can be obtained through the graph adjacency matrix A and the degree matrix D. T represents the transpose.

[0152] S302: Send the output of the second spectral graph wavelet packet convolutional layer into the graph readout layer, and the final representation of the graph is:

[0153]

[0154] where N is the number of graph nodes, is the final graph representation.

[0155] S303: Input the obtained fused graph representation features into the fully connected layer FCLayers for classification, and the prediction probability of a certain fault mode is:

[0156]

[0157] where softmax( ) can normalize the output of the fully connected layer into the probability distribution of the predicted output classes. FCLayers represents a single-layer fully connected layer,

[0158] S304: According to the obtained prediction probability of a certain fault mode, the cross-entropy loss function can be used to train the model, and it can be expressed as:

[0159]

[0160] Among them, M is the number of graphs, is the number of fault categories, is the sign function, which is 0 or 1. If the true category of the i-th sample is equal to take 1, otherwise take 0, is the predicted probability that the i-th sample belongs to the category .

[0161] In another embodiment, in step S301, the is the spectrogram wavelet packet transform operator, which can be obtained through the adjacency matrix A and degree matrix D of the graph. The steps are as follows:

[0162] S3011: First, obtain the Laplacian matrix L according to the adjacency matrix A and degree matrix D of the graph:

[0163]

[0164] S3012: Different from the spectrogram wavelet transform matrix that only uses the scaling function for low-pass filtering, any filter constructed by the spectrogram wavelet packet transform is the dilation product of the initial scaling function filter and the initial wavelet function filter . Its dilation can be expressed as , is the dilation scale factor. According to the spatial decomposition definition of the spectrogram wavelet packet, the first three-layer spectrogram wavelet packet operators can be obtained:

[0165]

[0166] The number of layers of the spectrogram wavelet packet can be continuously decomposed as the dilation factor dilates. Only the first three layers are shown here.

[0167] S3013: Since when calculating the scale kernel function and the wavelet kernel function , it is necessary to calculate the corresponding and , it is necessary to perform matrix decomposition on the Laplacian matrix to obtain the eigenvalues , which will lead to high computational complexity. To obtain better computational efficiency, the scale kernel function and the wavelet kernel function approximate by Chebyshev polynomials can be used:

[0168]

[0169] Among them, are the eigenvalues obtained by the eigenvalue decomposition of the Laplacian matrix L, , is the coefficient of the 0th term of and , , is the coefficient of the kth term of and , is the shifted Chebyshev polynomial with the domain , and its iterative form is:

[0170]

[0171] where is the largest eigenvalue, which can be quickly obtained by the Arnoldi algorithm without performing a complete matrix factorization. Therefore, the coefficient of the kth term of the shifted Chebyshev polynomial:

[0172]

[0173] where is the independent variable of the function, is the pi.

[0174] S3014: After obtaining the Chebyshev polynomial approximation coefficients of the scale kernel function and the wavelet kernel function at each dilation coefficient of the Chebyshev polynomial approximation, the Chebyshev polynomial coefficients after multiplying two Chebyshev polynomials can be obtained according to the following formula. Suppose there are two polynomials in Chebyshev form:

[0175]

[0176] where . The product of the polynomials has the following Chebyshev form:

[0177]

[0178] The Chebyshev coefficients of the new product polynomial can be obtained by the following formula:

[0179]

[0180] Accordingly, the spectral graph wavelet packet transform operator for the first layer and subsequent layers can be obtained and applied to the spectral graph wavelet packet convolutional neural network.

[0181] In another embodiment, in step S303, the fully connected layer FCLayers is expressed as follows:

[0182] ,

[0183] wherein, is the weight matrix to be learned, is the bias vector.

[0184] In another embodiment, in step S400, after each pair of models uses all the training set data to train and update the weights, the test set data is used to test the diagnostic accuracy of the model, and the test accuracy is obtained. It should be noted that the test accuracy is defined as: the number of samples correctly classified on the test set divided by the total number of samples. The model undergoes multiple rounds of training and learning, and the model with the highest test accuracy is saved as the variable working condition fault diagnosis model.

[0185] The following combines Figures 1 to 3 to further describe the technical solution of the present disclosure.

[0186] In a specific embodiment, as Figure 2 shown, when collecting multi-sensor signals of the gear fault of the drive system under time-varying speed and time-varying load, 4 vibration sensors, 2 current sensors, and 2 rotational speed signals and 2 torque signals collected by 2 encoders are used to simulate the operating states of the gearbox of the drive system under normal conditions, planet gear end breakage fault, planet gear spalling fault, and compound fault of planet gear crack and sun gear crack. The data sampling frequency is 10Khz. Among them, the time-varying speed means that the speed gradually increases from 500 rpm to 1300 rpm within 2 minutes, and the time-varying load means that the load increases from 10 NM to 30 NM within 2 minutes.

[0187] The signals of 10 of these sensors are used to construct multi-sensor graph signals. Each sensor intercepts every 1024 points as a node signal, and each graph has 10 nodes for 10 sensors. The node adjacency matrix is obtained through the RadiusGraph method, thereby constructing a multi-sensor graph signal data set. Among them, there are 1000 samples for each of the 4 states, with a total of 4000 samples. The samples are divided according to the ratio of training set:test set = 1:1. Therefore, the number of training set samples and the number of test set samples are both 2000.

[0188] Then, the training set data is input into the model of this method and the comparison model for learning and training. The model SGWPNet of this method uses Meyer as the kernel function, the decomposition layer number is 1 layer, and the Chebyshev approximation order is 10. The comparison model uses GCN and Graph Transformer. In addition, the experimental performance of the model under different signal-to-noise ratio (SNR) conditions is also simulated. Each group of experiments is trained for 100 epochs, and each model is tested 5 times respectively. The model accuracy is the average of 5 tests to eliminate the influence of test randomness.

[0189] Figure 3 It is the diagnostic result after inputting the test set data into the model. It can be seen from the figure that the model SGWPNet of this method can achieve the best accuracy under all SNR conditions, indicating that the anti-noise performance of the model is stable. While the diagnostic ability of GCN is poor, and the diagnostic ability of Graph Transformer is not as good as SGWPNet, and its diagnostic accuracy is also unstable with the change of noise. The above conclusions prove that the model of this method can effectively conduct fault diagnosis on the variable working condition data under the conditions of time-varying speed and time-varying load of the transmission system gearbox. By using spectral graph wavelet packet convolution to filter the refined graph spectrum and mining the fusion invariant features of time-varying graph signals, the fault diagnosis ability of the network under time-varying working condition conditions is enhanced.

[0190] Although the embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields. The above specific embodiments are merely illustrative and guiding, rather than restrictive. Those of ordinary skill in the art can also make many forms under the inspiration of this specification and without departing from the scope protected by the claims of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A transmission system fault diagnosis method, characterized in that: The steps include: Step S100: collecting multi-sensor signals of the transmission system when a gear fails with time-varying speed and time-varying load; Step S200: construct a multi-sensor network time-varying graph dataset and divide it into a training set and a test set; Step S300: input the training set data into the spectral wavelet packet convolutional neural network for training to achieve fault classification; Step S400: Use the test set data to test the trained spectral wavelet packet convolutional neural network, and save the spectral wavelet packet convolutional neural network with the best test performance for transmission system fault diagnosis.

2. A transmission system fault diagnosis method according to claim 1, characterized in that: Preferably, in step S100, the multi-sensor signal includes vibration, rotation speed, torque and current signals.

3. A transmission system fault diagnosis method according to claim 1, characterized in that: In step S200, constructing a multi-sensor network time-varying graph dataset includes: S201: For a multivariate time series X=[X1, X2, …, X τ ], for a single sensor X i The univariate sequence is divided into sub-samples of length l and a corresponding label is assigned to each sub-sample. There is no overlap between samples. The obtained sub-sample set is expressed as: ， In the formula, ∏ is the obtained subsample set, x represents the subsample, y represents the label, and n represents the number of subsamples; S202: According to the obtained sub-sample set Ώ, each sub-sample is regarded as a node of the multi-sensor graph network to construct a graph with τ nodes, and the sensor neighbors are found by calculating the cosine distance. The adjacency matrix is ​​obtained by the RadiusGraph method, including: ， Where -radius(·) is the calculation node x i and node x j Cosine similarity of , A(i,j) is the value of the i-th row and j-th column in the adjacency matrix of the multi-sensor graph, To determine whether there is an edge between nodes, the threshold is set to 0, and the node feature matrix V is obtained by dividing the sub-sample set Obtained by node correspondence.

4. A transmission system fault diagnosis method according to claim 1, characterized in that: In step S300, the spectral wavelet packet convolutional neural network is obtained by superimposing two spectral wavelet packet convolutional layers to extract multi-scale features of the graph spectrum and two fully connected layers to perform fault classification, which includes the following steps: S301: The two spectral wavelet convolutional layers of the spectral wavelet packet convolutional neural network are defined as follows: ， Among them, V and H represent the input graph node feature vector and the learned node representation respectively. and are the learnable filter coefficient matrices in the two spectral wavelet convolution layers, Relu() is a nonlinear activation function, is the spectral graph wavelet packet transform operator, which is obtained through the graph adjacency matrix A and the degree matrix D. T represents the transpose. S302: Send the output of the second spectral wavelet packet convolution layer to the graph readout layer, and the final graph representation is: ， Where N is the number of graph nodes, For the final graph representation, S303: Input the obtained fusion graph representation features into the fully connected layer FCLayers for classification. The predicted probability of the fault mode is: ， Among them, softmax() normalizes the output of the fully connected layer to the probability distribution of the predicted output class, FCLayers represents a single-layer fully connected layer, S304: According to the obtained failure mode prediction probability, the model is trained using a cross entropy loss function, which is expressed as: ， Where M is the number of graphs, is the number of fault categories, is a sign function, which is 0 or 1 if the true category of the i-th sample is equal to Take 1, otherwise take 0. To observe that the i-th sample belongs to the category The predicted probability of .

5. A transmission system fault diagnosis method according to claim 4, characterized in that: In step S301, the spectral wavelet packet transform operator It is obtained through the adjacency matrix A and degree matrix D of the graph. The steps are as follows: S3011: First, obtain the Laplace matrix L according to the adjacency matrix A and degree matrix D of the graph: ， S3012: Any filter constructed by spectral wavelet packet transform is an initial scale function filter With the initial wavelet function filter The expansion product of , is the expansion scale factor. According to the spatial decomposition definition of the spectral wavelet packet, the first three layers of the spectral wavelet packet operators are obtained: ， The number of layers of the spectral wavelet packet continues to decompose as the expansion factor expands. Only the first three layers are shown here. S3013: Scaling kernel function using Chebyshev polynomial approximation and wavelet kernel function : ， in, is the eigenvalue obtained by decomposing the Laplace matrix eigenvalue L, , is represented by Chebyshev polynomials and The 0th coefficient of , is represented by Chebyshev polynomials and The kth coefficient of The domain is The shifted Chebyshev polynomial, its iterative form is: ， in, is the maximum eigenvalue, obtained by the Arnoldi algorithm, the kth coefficient of the shifted Chebyshev polynomial: ， in, is the function variable, is the circumference of a circle, S3014: After obtaining the Chebyshev polynomial approximation coefficients of the scaling kernel function and the wavelet kernel function under each expansion coefficient of the Chebyshev polynomial approximation, the Chebyshev polynomial coefficients after multiplying two Chebyshev polynomials are obtained according to the following formula, two Chebyshev form polynomials: ， in, , the product of polynomials There are the following Chebyshev forms: ， Chebyshev coefficients of the new product polynomial Obtained by the following formula: ， Based on this, we can obtain the first layer and subsequent spectral wavelet packet transform operators , and applied it to the spectral wavelet packet convolutional neural network.

6. A transmission system fault diagnosis method according to claim 4, characterized in that: In step S303, the fully connected layer FCLayers is represented as follows: ， in, is the weight matrix to be learned, is the bias vector.

7. A transmission system fault diagnosis method according to claim 1, characterized in that: In step S400, after each pair of spectral wavelet packet convolutional neural networks is trained and updated with all the training set data, the diagnostic accuracy of the spectral wavelet packet convolutional neural network is tested using the test set data to obtain the test accuracy of the number of correctly classified samples on the test set divided by the total number of samples. The model is trained and learned in multiple cycles, and the spectral wavelet packet convolutional neural network with the highest test accuracy is saved as the spectral wavelet packet convolutional neural network for variable operating condition fault diagnosis.

8. A transmission system fault diagnosis system, characterized in that: These include, A collection unit for collecting multi-sensor signals of a transmission system when a gear fails with a time-varying speed and a time-varying load; A construction unit, which is used to construct a multi-sensor network time-varying graph dataset and divide it into a training set and a test set; A training unit, which is used to input the training set data into the spectral wavelet packet convolutional neural network for training to achieve fault classification; The testing unit is used to test the trained spectral wavelet packet convolutional neural network using the test set data, and save the spectral wavelet packet convolutional neural network with the best test performance for transmission system fault diagnosis.

9. A computer storage medium, characterized in that The storage medium includes computer instructions, which, when executed on a computer, enable the computer to execute the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The electronic device comprises: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

Images