Planetary gearbox fault diagnosis method based on graph convolution pulse residual network

Abstract

The invention relates to the technical field of gear fault diagnosis, in particular to a planetary gearbox fault diagnosis method based on a graph convolution pulse residual network. According to the fault diagnosis method based on the graph convolution pulse residual network, operation information in different fault states is collected, graph data is constructed through data processing, the exclusive model is trained, and accurate prediction of fault types is achieved. The method has the advantages that the graph convolution network and the pulse residual network are combined, the capturing capability of graph convolution on the data dependence relation is reserved, the gradient problem of a deep network is relieved by means of a residual structure, meanwhile, pulse neuron characteristics are integrated, the recognition precision and network stability of fault features under the complex working condition are improved, and the fault recognition efficiency is improved. And an efficient and reliable brand-new way is provided for planetary gearbox fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of gear fault diagnosis technology, specifically a planetary gearbox fault diagnosis method based on graph convolutional pulse residual networks. Background Technology

[0002] Planetary gearboxes, with their significant advantages such as large transmission ratio, stable power transmission, and strong load-bearing capacity, are widely used in large, low-speed, heavy-duty equipment such as coal mine roadway tunneling machines, coal mining machines, port cranes, heavy trucks, and tunnel boring machines. However, industrial operating environments are often harsh. As a core component of the power transmission system, planetary gearboxes operate under complex conditions such as high load and low speed for extended periods, which can easily lead to failures in critical components such as bearings and gears. This can result in damage to the equipment's power system and reduced production efficiency, or even major safety accidents. Therefore, research on fault diagnosis technology for low-speed, heavy-duty planetary gearboxes is of great significance for reducing industrial equipment maintenance costs, minimizing unplanned downtime, and improving production efficiency.

[0003] Fault diagnosis methods for planetary gearboxes are mainly divided into two categories: feature extraction-based fault diagnosis methods and data-driven fault diagnosis methods. However, feature extraction-based fault diagnosis methods rely heavily on domain knowledge and expert experience, and are easily affected by subjective factors. Furthermore, with the explosive growth of industrial field data, traditional feature extraction methods struggle to efficiently process massive amounts of data. In contrast, data-driven fault diagnosis methods can automatically extract effective information from massive amounts of raw data, and have become an important means of identifying planetary gearbox fault conditions in the industrial big data environment.

[0004] In the context of industrial big data, deep learning methods such as convolutional neural networks (CNNs), long short-term memory (LSTM) neural networks, and recurrent neural networks (RNNs) have effectively solved the problems of surging data volume and automatic data feature extraction, achieving certain application results in the field of planetary gearbox fault diagnosis. However, these neural networks can only capture the hidden features of conventional data (such as images and time series), ignoring the interdependencies between data. When identifying weak fault features in a noisy environment, their diagnostic accuracy is low. To address this issue, graph convolutional neural networks (GNNs) have been introduced into the field of planetary gearbox fault diagnosis, effectively improving diagnostic accuracy. However, as a second-generation neural network, it focuses on mining spatial features of signals but neglects the representation of temporal information, resulting in a lack of biological interpretability as an artificial intelligence method. Furthermore, existing GNNs generally have a relatively small number of layers, making them prone to gradient vanishing problems when expanding depth, which severely limits their application in planetary gearbox fault diagnosis tasks under noisy conditions. Summary of the Invention

[0005] To avoid and overcome the technical problems existing in the prior art, this invention provides a planetary gearbox fault diagnosis method based on a graph convolutional pulse residual network. This invention combines a graph convolutional network with a pulse residual network, retaining the ability of graph convolution to capture data dependencies, alleviating the gradient problem in deep networks by leveraging the residual structure, and incorporating the characteristics of pulse neurons to improve the accuracy of fault feature recognition and network stability under complex operating conditions, thus providing a new, efficient, and reliable approach for planetary gearbox fault diagnosis.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A fault diagnosis method for planetary gearboxes based on graph convolutional pulse residual networks includes the following diagnostic steps:

[0008] S1. Collect operating information of key components of the planetary gearbox under different fault states under stable operating conditions;

[0009] S2. Normalize the running information to obtain normalized running information; use window sliding and Fourier transform to process the normalized running information to obtain a frequency domain dataset.

[0010] S3. Use the K-nearest neighbor method to construct the frequency domain dataset into graph data;

[0011] S4. Train the graph convolutional pulse residual network using graph data to obtain a fault diagnosis model;

[0012] S5. Predict the failure types of key components of the planetary gearbox through a fault diagnosis model.

[0013] As a further aspect of the present invention: Operational information X = [x1, ..., x i ,…,x K ], x i =[x i (1),…,x i (n),…x i (M)],x i Let x represent the i-th fault state. i (n) represents the nth sampled data of the i-th fault state, i∈[1,K], K is the total number of fault state types, n∈[1,M], M is the sampled signal x i (n) is the number of sampling points contained in the sample.

[0014] As a further aspect of the present invention, the process of obtaining the frequency domain dataset is as follows:

[0015] S21. Normalize the running information X to obtain normalized running information.

[0016]

[0017] In the formula, x represents i Normalized data; x represents i (n) Normalized data; min(x) i ) represents x i The minimum value in; max(x) i ) represents x i The maximum value in;

[0018] S22. Set the window length to L, and use a sliding window non-overlapping resampling method to normalize the running information. The data is segmented to obtain a dataset consisting of multiple sets of data samples of length L:

[0019]

[0020] In the formula, m represents the number of window slides; floor(·) represents the floor function.

[0021] S23. Perform a Fourier transform on the dataset 'data' to obtain the frequency domain dataset 'DATA':

[0022] DATA = FFT(data);

[0023] In the formula, FFT represents Fourier transform.

[0024] As a further aspect of the present invention, the process of constructing graph data is as follows:

[0025] S31. Based on a graph convolutional network structure, each data sample under the same fault state is treated as a node in the graph, and the nodes are connected to form graph data; the K-nearest neighbor method is used to find the first s nearest neighbors of each node to obtain node x. fu The nearest neighbors are shown in the following formula:

[0026] Ne(x fu ) = KNN(s,x fu ,ψ)

[0027] Wherein, KNN(·) represents the u-th node x in the frequency domain dataset DATA. fu The first s nearest neighbors; Ψ = [x f1 ,…,x fu ,…,x fv ,…,x fl ], Ψ represents a subset containing x nodes, x fv Let u, v ∈ [1, l] represent the v-th node; Ne(x) fu ) represents node xfi The nearest neighbor;

[0028] S32. The edge weights between connected nodes are calculated using the Gaussian kernel weighting function, as shown in the following formula:

[0029]

[0030] Among them, e uv Represents node x fu and node x fv Edge weights between ||(x) fu ,x fv )|| 2 Represents node x fu and node x fv The Euclidean distance between them; ζ is the bandwidth of the Gaussian kernel, which is the average value of all Euclidean distances;

[0031] S33. Divide the constructed graph data into training set and test set according to the set ratio.

[0032] As a further aspect of the present invention, the training process of the diagnostic model is as follows:

[0033] S41. Construct a graph convolutional spiking residual network. The first layer consists of a graph convolutional layer, a batch normalization layer, and a ReLU activation function layer. The second to fourth layers are graph convolutional spiking residual blocks. The fifth layer consists of a fully connected layer based on spiking neurons and a softmax function layer.

[0034] S42. Initialize the network hyperparameters, input the training set into the graph convolutional pulse residual network in batches for feature extraction, and train the network structure using the cross-entropy loss function according to the data labels.

[0035] S43. Repeat step S42 until the preset maximum number of iterations is reached to obtain the trained graph convolutional pulse residual network, which is the fault diagnosis model.

[0036] As a further aspect of the present invention, the cross-entropy loss function is expressed as follows:

[0037] L CE (y label ,y predicted )=-∑y label log(y predicted );

[0038] In the formula, y predicted and y label These are the predicted labels and the actual labels, respectively.

[0039] As a further aspect of the present invention: after obtaining the fault diagnosis model, the test set is input into the fault diagnosis model to predict and classify different fault states of key components of the planetary gearbox.

[0040] As a further embodiment of the present invention: the graph convolutional pulse residual block is composed of graph convolutional layers, batch normalization layers and integral ignition pulse neuron layers.

[0041] As a further aspect of the present invention: the fully connected layer based on spiking neurons is constructed from leaky integral ignition spiking neurons, and its expression is as follows:

[0042]

[0043] In the formula, τ m Let ω represent the capacitance constant, u(t) represent the membrane potential of the neuron at time t, I(t) represent the input current at time t, and ω represent the capacitance constant. r and s r (t) represents the synaptic weight and output pulse of the r-th neuron in the last layer at time t, where r∈[1,R], R represents the total number of neurons in the last layer; θ represents the activation threshold of the neuron; u rest This represents the resting potential of a neuron.

[0044] 10. A planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 9, wherein the different fault states of the key components of the planetary gearbox include normal, crack, spalling, broken teeth and missing teeth.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. This invention presents a fault diagnosis method based on graph convolutional pulsed residual networks. By collecting operational information under different fault states, processing the data to construct graph data, and training a dedicated model, it achieves accurate prediction of fault types. Its advantages lie in combining graph convolutional networks with pulsed residual networks. This retains the ability of graph convolution to capture data dependencies while mitigating the gradient problem of deep networks through residual structures. Furthermore, the integration of pulsed neuron characteristics improves the accuracy of fault feature recognition and network stability under complex operating conditions, providing an efficient and reliable new approach for planetary gearbox fault diagnosis.

[0047] 2. By clearly defining the symbols and meanings of different fault states and corresponding sampled data, the data description becomes more standardized and logical. This advantage facilitates the accurate execution of subsequent data processing steps (such as normalization and Fourier transform), ensuring the consistency of data transmission throughout the diagnostic process and laying a solid foundation for building high-quality datasets and training models. First, normalization eliminates interference from data of different magnitudes, ensuring data comparability. Then, sliding window non-overlapping resampling is used for data segmentation, which can more comprehensively capture the local features of the data. Finally, Fourier transform converts the time-domain signal to the frequency-domain signal, effectively extracting frequency features from the signal, facilitating subsequent analysis and identification of fault features, and significantly improving the quality and effectiveness of the dataset.

[0048] 3. Using the K-nearest neighbor method to determine the nearest neighbor relationship of nodes can accurately uncover the inherent correlation between data samples, making the constructed graph data more consistent with the actual data distribution; using the Gaussian kernel weighting function to calculate the edge weights can assign different weights according to the Euclidean distance between nodes, enhancing the graph data's ability to represent sample differences; and dividing the training set and test set proportionally can ensure the objectivity of model training and evaluation, improving the model's generalization ability.

[0049] 4. The constructed graph convolutional spiking residual network has a clear hierarchical structure. The first layer, consisting of a graph convolutional layer, a batch normalization layer, and a ReLU activation function layer, effectively extracts initial features and accelerates training. The middle graph convolutional spiking residual block deepens the network, enhances feature extraction capabilities, and alleviates the gradient vanishing problem. Finally, the fully connected layer based on spiking neurons and the softmax function layer achieve accurate classification of fault types. Training with the cross-entropy loss function effectively measures the difference between predicted and true labels, ensuring the accuracy and efficiency of model training. The penalty for prediction errors increases with the severity of the error, accurately reflecting the difference between predicted and true labels. This effectively guides the optimization of network parameters, enabling the model to converge to its optimal state faster during training and improving the model's diagnostic accuracy.

[0050] 5. The graph convolutional spiking residual block consists of graph convolutional layers, batch normalization layers, and integral ignition spiking neuron layers. Its advantages lie in the graph convolutional layers extracting spatial features, the batch normalization layers accelerating network training and improving stability, and the integral ignition spiking neuron layers introducing spiking neural characteristics to better capture temporal information and dynamic features. The combination of these three significantly enhances the network's ability to extract and represent complex fault features. The fully connected layer constructed based on leaky integral ignition spiking neurons considers factors such as capacitance constant, membrane potential, input current, synaptic weights, activation threshold, and resting potential, more closely resembling the working mechanism of biological neurons and enhancing the network's biological interpretability. Simultaneously, this structure effectively processes time-series information, improving the model's ability to identify dynamic fault features and making fault diagnosis more accurate.

[0051] 6. Different fault states of key components of planetary gearboxes, including normal, crack, spalling, broken teeth and missing teeth, cover common fault types of planetary gearboxes. Its advantage is that it makes the target of fault diagnosis clearer and more specific, which facilitates targeted training and identification of the model, improves the practicality and pertinence of the diagnostic method, and can better meet the needs of planetary gearbox fault diagnosis in actual industrial scenarios. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the diagnostic process of the present invention.

[0053] Figure 2 This is a schematic diagram of the planetary gearbox fault simulation test bench used in this embodiment of the invention.

[0054] Figure 3 This is a time-domain waveform diagram of the vibration signal collected under normal conditions in an embodiment of the present invention.

[0055] Figure 4 This is a time-domain waveform diagram of the vibration signal collected during the peeling of the sun gear in an embodiment of the present invention.

[0056] Figure 5 This is a time-domain waveform of the vibration signal collected in an embodiment of the present invention when the sun gear has a 1mm crack.

[0057] Figure 6 This is a time-domain waveform of the vibration signal collected in an embodiment of the present invention when the sun gear has a crack of 1.8 mm.

[0058] Figure 7 This is a time-domain waveform of the vibration signal collected in an embodiment of the present invention when the sun gear has a crack of 2.5 mm.

[0059] Figure 8 This is a time-domain waveform diagram of the vibration signal collected in an embodiment of the present invention when the sun gear tooth breaks.

[0060] Figure 9 This is a time-domain waveform diagram of the vibration signal collected in an embodiment of the present invention when the sun gear has a missing tooth.

[0061] Figure 10 This is a time-domain waveform diagram of the vibration signal collected during the peeling of the planetary gear in an embodiment of the present invention.

[0062] Figure 11 This is a time-domain waveform of the vibration signal collected in an embodiment of the present invention when the planetary gear has a 1mm crack.

[0063] Figure 12 This is a time-domain waveform of the vibration signal collected in an embodiment of the present invention when the planetary gear has a crack of 1.8 mm.

[0064] Figure 13 This is a time-domain waveform of the vibration signal collected in an embodiment of the present invention when the planetary gear has a crack of 2.5 mm.

[0065] Figure 14 This is a time-domain waveform diagram of the vibration signal collected in an embodiment of the present invention when a planetary gear tooth breaks.

[0066] Figure 15 This is a time-domain waveform diagram of the vibration signal of a planetary gear with missing teeth, collected in an embodiment of the present invention.

[0067] Figure 16 This is a time-domain waveform diagram of the vibration signal collected in an embodiment of the present invention when the inner ring of a planetary bearing is faulty.

[0068] Figure 17 This is a time-domain waveform diagram of the vibration signal collected in an embodiment of the present invention when the outer ring of a planetary bearing fails.

[0069] Figure 18 This is a time-domain waveform diagram of the vibration signal collected in an embodiment of the present invention when the sun gear and planetary gear are cracked.

[0070] Figure 19 This is a time-domain waveform diagram of the vibration signal collected in an embodiment of the present invention when the sun gear peels off and the planetary gear cracks.

[0071] Figure 20 This is a flowchart illustrating the construction process of the fault diagnosis model in an embodiment of the present invention.

[0072] Figure 21 This is a matrix diagram of diagnostic results in an embodiment of the present invention.

[0073] In the diagram: 1. Mounting base; 2. Three-phase drive motor; 3. Coupling; 4. Planetary reduction gearbox; 5. Torque sensor I; 6. Planetary speed increase gearbox; 7. Torque sensor II; 8. Three-phase load motor; 9. Vibration damping base. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0075] Please see Figure 1 In this embodiment of the invention, a planetary gearbox fault diagnosis method based on graph convolutional pulse residual network includes the following:

[0076] I. Obtaining Operational Information

[0077] The planetary gearbox fault simulation test bench used in this embodiment is as follows: Figure 2 As shown in the figure, the structure of the planetary gearbox fault simulation test bench used in this embodiment is as follows: Figure 2 As shown, the test bench is assembled from core components such as mounting base 1, three-phase drive motor 2, coupling 3, planetary reduction gearbox 4, torque sensor I5, planetary speed-increasing gearbox 6, torque sensor II7, three-phase load motor 8, and shock-absorbing base 9.

[0078] The specific connections are as follows: the three-phase drive motor 2 is fixedly mounted on the mounting base 1, and its output end is connected to the planetary reduction gearbox 4 via a coupling 3; the output end of the planetary reduction gearbox 4 and the input end of the planetary speed-increasing gearbox 6 are connected coaxially via a torque sensor I5; the output end of the planetary speed-increasing gearbox 6 is connected coaxially to the input end of the three-phase load motor 8 via a torque sensor II7. All of the above devices are assembled on the vibration-damping base 9 to ensure the stability of the test bench operation.

[0079] In this embodiment, the speed of the three-phase drive motor 2 is set to 150 r / min, and the three-phase load motor 8 is adjusted to provide a load torque of 300 N·m through torque sensor I5 and torque sensor II7. The basic parameters of the planetary reduction gearbox 4 are shown in Table 1.

[0080] Table 1 Basic parameters of planetary reduction gearbox

[0081]

[0082] An accelerometer was attached to the surface of the planetary gearbox 4 housing at the position corresponding to the gear ring. The sampling frequency was set to 12800Hz and the sampling duration to 300s. Operating information of key components of the planetary gearbox 4 under different fault conditions was collected during stable operation at 150r / min. The time-domain waveform of the collected vibration signal is shown below. Figure 3-19As shown, there are a total of seventeen fault states.

[0083] II. Obtaining the Frequency Domain Dataset

[0084] First, the runtime information X is normalized to obtain normalized runtime information.

[0085]

[0086]

[0087] In the formula, x represents i Normalized data; x represents i (n) Normalized data; min(x) i ) represents x i The minimum value in; max(x) i ) represents x i The maximum value in.

[0088] Next, the window length was set to L = 12800, and a sliding window non-overlapping resampling method was used to normalize the running information. The data is segmented to obtain a dataset consisting of 5100 data samples of length L:

[0089]

[0090] In the formula, m represents the number of window slides; floor(·) represents the floor function.

[0091] Finally, perform a Fourier transform on the dataset 'data' to obtain the frequency domain dataset 'DATA'.

[0092] DATA = FFT(data);

[0093] In the formula, FFT represents Fourier transform.

[0094] III. Constructing Graph Data

[0095] First, based on the graph convolutional network structure, each data sample under the same fault state is treated as a node in the graph, and the nodes are connected to form graph data. The K-nearest neighbor method is used to find the first s nearest neighbors of each node to obtain node x. fu The nearest neighbors are shown in the following formula:

[0096] Ne(x fu ) = KNN(s,x fu ,Ψ)

[0097] Wherein, KNN(·) represents the u-th node x in the frequency domain dataset DATA. fu The first s nearest neighbors; Ψ = [x f1 ,…,x fu ,…,x fv ,…,x fl ], where ψ represents a subset containing l nodes, x fv Let u, v ∈ [1, l] represent the v-th node; Ne(x) fu ) represents node x fi . the neighbor.

[0098] Next, the edge weights between connected nodes are calculated using a Gaussian kernel weighting function, as shown in the following formula:

[0099]

[0100] Among them, e uv Represents node x fu and node x fv Edge weights between; ||(x fu ,x fv )‖ 2 Represents node x fu and node x fv The Euclidean distance between them; ζ is the bandwidth of the Gaussian kernel, which is the average value of all Euclidean distances;

[0101] Finally, the constructed graph data was divided into a training set (80%) and a test set (20%) according to the set ratio. The training set contained 4080 data samples, and the test set contained 1020 data samples, as detailed in Table 2.

[0102] Table 2 Dataset Distribution Details

[0103]

[0104]

[0105] IV. Obtaining the Fault Diagnosis Model

[0106] like Figure 20 As shown, firstly, a graph convolutional spiking residual network is constructed. The first layer consists of a graph convolutional layer, a batch normalization layer, and a ReLU activation function layer. The second to fourth layers are graph convolutional spiking residual blocks. The fifth layer consists of a fully connected layer based on spiking neurons and a softmax function layer.

[0107] Then, the network hyperparameters are initialized as shown in Table 3. The training set is then batch-input into the graph convolutional pulse residual network for feature extraction. Based on the data labels, the network structure is trained using the cross-entropy loss function. The cross-entropy loss function is expressed as follows:

[0108] L CE (y label ,y predicted )=-∑y label log(y predicted );

[0109] In the formula, y predicted and y label These are the predicted labels and the actual labels, respectively.

[0110] Table 3 shows the network hyperparameters in the examples.

[0111]

[0112] Finally, the training is repeated until the preset maximum number of iterations is reached, resulting in a trained graph convolutional pulse residual network, which is the fault diagnosis model.

[0113] V. Predicting Fault Types

[0114] The test set is input into the fault diagnosis model to classify different fault states of key components of the planetary gearbox, diagnose fault types, and the diagnosis results are as follows: Figure 21 As shown. Figure 5 In the confusion matrix shown, only 11 data samples labeled 14 were incorrectly diagnosed as labeled 9, meaning that 11 samples of planetary bearing outer ring faults were incorrectly diagnosed as planetary gear cracks of 1.8mm. The calculated diagnostic accuracy was 98.92%. The diagnostic results indicate that the planetary gearbox fault diagnosis method based on graph convolutional pulse residual network provided by this invention can effectively achieve accurate diagnosis of different fault states of planetary gearboxes under conditions of weak fault characteristics, strong noise interference, and industrial big data.

[0115] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of planetary gearboxes based on graph convolutional pulse residual networks, characterized in that, The diagnostic steps include the following: S1. Collect operating information of key components of the planetary gearbox under different fault states under stable operating conditions; S2. Normalize the operation information to obtain normalized operation information; The normalized runtime information is processed using window sliding and Fourier transform to obtain a frequency domain dataset; S3. Use the K-nearest neighbor method to construct the frequency domain dataset into graph data; S4. Train the graph convolutional pulse residual network using graph data to obtain a fault diagnosis model; S5. Predict the failure types of key components of the planetary gearbox through a fault diagnosis model.

2. The planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 1, characterized in that, Running information X = [x1,..., x i ,..., x K ], x i = [x i (1),..., x i (n),... x i (M)], x i represents the i-th fault state, x i (n) represents the n-th sampling data of the i-th fault state, i ∈ [1, K], K is the total number of fault state types, n ∈ [1, M], M is the number of sampling points contained in the collected signal x i (n).

3. The planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 2, characterized in that, The process of obtaining the frequency domain dataset is as follows: S21. Normalize the running information X to obtain normalized running information. In the formula, x represents i Normalized data; x represents i (n) Normalized data; min(x) i ) represents x i The minimum value in; max(x) i ) represents x i The maximum value in; S22. Set the window length to L, and use a sliding window non-overlapping resampling method to normalize the running information. The data is segmented to obtain a dataset consisting of multiple sets of data samples of length L: In the formula, m represents the number of window slides; floor(·) represents the floor function. S23. Perform a Fourier transform on the dataset 'data' to obtain the frequency domain dataset 'DATA': DATA = FFT(data); In the formula, FFT represents Fourier transform.

4. The planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 3, characterized in that, The process of constructing graph data is as follows: S31, based on the graph convolution network structure, each data sample under the same fault state is taken as a node in the graph, and the nodes are connected to form graph data; the K nearest neighbor method is used to find the first s nearest neighbors of each node, and the nearest neighbors of the node x fu are as follows: Ne(x fu ) = KNN(s, x fu , Ψ) Wherein, KNN(·) represents the u-th node x in the frequency domain dataset DATA. fu The first s nearest neighbors; Ψ = [x f1 ,…,x fu ,…,x fv ,…,x fl ], Ψ represents a subset containing l nodes, x fv Let u, v ∈ [1, l] represent the v-th node; Ne(x) fu ) represents node x fi The nearest neighbor; S32. The edge weights between connected nodes are calculated using the Gaussian kernel weighting function, as shown in the following formula: Among them, e uv Represents node x fu and node x fv Edge weights between ||(x) fu ,x fv )|| 2 Represents node x fu and node x fv The Euclidean distance between them; ζ is the bandwidth of the Gaussian kernel, which is the average value of all Euclidean distances; S33. Divide the constructed graph data into training set and test set according to the set ratio.

5. The planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 4, characterized in that, The training process of the diagnostic model is as follows: S41. Construct a graph convolutional spiking residual network. The first layer consists of a graph convolutional layer, a batch normalization layer, and a ReLU activation function layer. The second to fourth layers are graph convolutional spiking residual blocks. The fifth layer consists of a fully connected layer based on spiking neurons and a softmax function layer. S42. Initialize the network hyperparameters, input the training set into the graph convolutional pulse residual network in batches for feature extraction, and train the network structure using the cross-entropy loss function according to the data labels. S43. Repeat step S42 until the preset maximum number of iterations is reached to obtain the trained graph convolutional pulse residual network, which is the fault diagnosis model.

6. The planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 5, characterized in that, The cross-entropy loss function is expressed as follows: L CE (and label ,and predicted )=-∑y label log(y predicted ); In the formula, y predicted and y label These are the predicted labels and the actual labels, respectively.

7. The planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 6, characterized in that, After obtaining the fault diagnosis model, the test set is input into the fault diagnosis model to predict and classify different fault states of key components of the planetary gearbox.

8. The planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 7, characterized in that, The graph convolutional pulse residual block consists of graph convolutional layers, batch normalization layers, and integral ignition pulse neuron layers.

9. A planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 8, characterized in that, The fully connected layer based on spiking neurons is constructed from leaky integral ignition spiking neurons, and its expression is as follows: In the formula, τ m Let ω represent the capacitance constant, u(t) represent the membrane potential of the neuron at time t, I(t) represent the input current at time t, and ω represent the capacitance constant. r and s r (t) represents the synaptic weight and output pulse of the r-th neuron in the last layer at time t, where r∈[1,R], R represents the total number of neurons in the last layer; θ represents the activation threshold of the neuron; u rest This represents the resting potential of a neuron.

10. A planetary gearbox fault diagnosis method based on graph convolutional pulse residual network according to claim 9, characterized in that, The different fault states of the key components of the planetary gearbox include normal, crack, peeling, broken teeth, and missing teeth.

Images