Gearbox fault diagnosis method and device based on mutual information entropy graph

Abstract

The present invention discloses a gearbox fault diagnosis method and device based on mutual information entropy graph construction, belonging to the field of gearbox fault diagnosis. The method includes extracting time series nodes from gearbox vibration data under different states to construct time series node sets under each state; preprocessing the time series node sets under each state respectively to obtain the preprocessed time series node sets under each state; based on the time series dependency of the time series nodes, using the mutual information entropy graph construction method to mine the correlation relationship of the time series nodes for the preprocessed time series node sets under each state, and constructing the entropy node graph set accordingly; introducing LSTM to construct a memory graph neural network model, using the entropy node graph set as the memory graph neural network model input, and mining the spatiotemporal correlation information of the gearbox vibration data through training and learning; using the trained memory graph neural network model to diagnose the gearbox data to be diagnosed. The present invention can effectively diagnose gearbox faults.

Description

Technical Field

[0001] The invention relates to a gearbox fault diagnosis method based on a mutual information entropy graph, and belongs to the field of gearbox fault diagnosis. Background Art

[0002] Gearboxes are one of the most widely used equipment in industrial production. Due to their advantages such as high transmission efficiency, compact structure and long service life, gearboxes are widely used in various industrial fields, including important rotating mechanical equipment such as high-speed trains, aerospace, and power systems. The health of the gearbox is directly related to the performance, stability and life of the equipment. On the production line, any failure of any equipment may lead to the interruption of the entire production line. In order to prevent accidents, reduce economic losses, and even avoid irreversible life safety risks, research on gearbox fault diagnosis is particularly important.

[0003] In the fault diagnosis of industrial equipment, especially for complex mechanical systems such as gearboxes, accurately capturing the dependencies and timing characteristics between components is the key to achieving efficient and accurate diagnosis. However, traditional graph neural network models mainly deal with the relationship network and timing dependencies of such complex equipment by building correlations, which makes it difficult to fully explore the potential feature relationships in the data. At the same time, vibration data contains a lot of noise, which further increases the complexity of model training and reduces the reliability of diagnosis. For example, graph construction methods based on KNN and Radius rely on distance metrics and node correlation when constructing node connections, which makes it difficult to capture the nonlinear dependencies between nodes; these methods can only simply connect nodes based on adjacent distances, ignoring the potential complex feature interactions in vibration data, resulting in the model's performance in actual complex fault diagnosis tasks. In addition, conventional graph convolutional neural networks (GCNs) usually do not consider the timing characteristics of node data, while the vibration data of industrial equipment often has significant time dependencies. Therefore, when processing industrial data with strong time dependencies, existing methods often find it difficult to obtain ideal diagnostic effects, limiting the application scenarios and diagnostic accuracy of the models. Summary of the invention

[0004] The present invention provides a gearbox fault diagnosis method and device based on a mutual information entropy graph, which is used to construct an entropy node atlas based on the mutual information entropy graph, further implement the training and verification of a memory graph neural network model based on the entropy node atlas, and then diagnose the gearbox fault based on the trained memory graph neural network model.

[0005] The technical solution of the present invention is:

[0006] According to a first aspect of the present invention, a gearbox fault diagnosis method based on a mutual information entropy graph is provided, comprising:

[0007] Step 1: extracting time series nodes from the gearbox vibration data under different states to construct a time series node set under each state; preprocessing the time series node set under each state respectively to obtain the preprocessed time series node set under each state;

[0008] Step 2: Based on the temporal dependency of the temporal nodes, the mutual information entropy graph construction method is used to mine the correlation relationship of the temporal nodes for the preprocessed temporal node sets in each state, and an entropy node graph set is constructed accordingly;

[0009] Step 3: Introduce LSTM to build a memory graph neural network model, use the entropy node graph set as the input of the memory graph neural network model, and mine the spatiotemporal correlation information of the gearbox vibration data through training and learning to realize the training of the memory graph neural network model;

[0010] Step 4: Use the trained memory graph neural network model to diagnose the gearbox data to identify the gearbox fault state to achieve gearbox fault diagnosis.

[0011] Furthermore, the step 1 is specifically as follows: according to a preset segmented data length, the gearbox vibration data under different states are segmented to obtain segmented sub-sample data, which are used as time series nodes to construct a time series node set; the time series node set under each state is normalized and then Fourier transformed to obtain the pre-processed time series node set under each state.

[0012] Furthermore, the step 2 is specifically as follows:

[0013] According to the preset number of time series nodes, the preprocessed time series node set in each state is divided into nodes according to the time series to obtain multiple sub-time series node sets in each state;

[0014] For each sub-series node set in each state, the joint histogram between any two sequence nodes and the edge histogram of the corresponding two sequence nodes are calculated respectively;

[0015] The joint histogram between two time series nodes and the corresponding edge histograms of the two time series nodes are smoothed respectively, and the joint probability distribution and the edge probability distribution are calculated;

[0016] According to the joint probability distribution and the marginal probability distribution, the joint entropy and marginal entropy of the corresponding two time series nodes are calculated;

[0017] According to the joint entropy and edge entropy of two time series nodes, the mutual information entropy of two time series nodes is calculated;

[0018] The mutual information entropy is used as the weight of the edge between two time series nodes to construct the entropy node graph.

[0019] Furthermore, the memory graph neural network model includes:

[0020] Input layer, used to take entropy node graph as input;

[0021] The first graph convolution layer is used to extract graph features from the input entropy node graph set, and aggregate the feature information of each time series node and its neighboring time series nodes through the graph convolution operation to obtain the first graph feature;

[0022] The first BatchNorm layer is used to perform batch normalization on the first graph features to eliminate internal covariate shift;

[0023] The second graph convolution layer is used to extract graph features from the first graph features after batch normalization, further aggregate wider neighbor features through graph convolution, and obtain higher-order information to obtain the second graph features;

[0024] The second BatchNorm layer is used to perform batch normalization on the second graph features;

[0025] FC layer, based on the batch normalized second graph features, the fully connected layer is used to reduce the feature dimension and integrate the features to obtain the FC layer graph features;

[0026] The Dropout layer is used to randomly discard some neurons in the FC layer graph features according to the set drop probability to obtain the Dropout layer graph features;

[0027] LSTM layer, which is used to input the graph features of the Dropout layer as time series and expand the features into time steps. It dynamically updates the memory through the input gate, forget gate, and output gate mechanism, and processes the information of node features changing over time to fuse graph features and time series features.

[0028] The output layer is used to map the output of the LSTM layer and output the state category of the gearbox vibration data.

[0029] According to a second aspect of the present invention, a gearbox fault diagnosis device constructed based on a mutual information entropy graph is provided, comprising:

[0030] An acquisition module is used to extract time series nodes from the gearbox vibration data under different states to construct a time series node set under each state; preprocess the time series node set under each state respectively to obtain the preprocessed time series node set under each state;

[0031] A construction module is used to mine the correlation relationship of the time series nodes based on the time series dependency relationship of the time series nodes, use the mutual information entropy graph construction method to build the time series node set preprocessed in each state, and construct the entropy node graph set accordingly;

[0032] The training module is used to introduce LSTM to build a memory graph neural network model, use the entropy node graph set as the input of the memory graph neural network model, and mine the spatiotemporal correlation information of the gearbox vibration data through training learning to realize the training of the memory graph neural network model;

[0033] The diagnostic module is used to diagnose the gearbox data to be diagnosed using the trained memory graph neural network model, identify the gearbox fault state, and realize the gearbox fault diagnosis.

[0034] The beneficial effect of the present invention is that the present invention mines graph structure data through the mutual information entropy graph construction method, and then inputs the memory graph neural network to build a fault diagnosis model to evaluate the health status of the gearbox. Compared with the existing methods, this method can capture the potential dependencies between nodes and filter irrelevant noise, which can effectively enhance the robustness and effective diagnostics of the model, and provide new technical routes and implementation methods for researchers in the field of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the present invention;

[0036] Figure 2 It is a flow chart for constructing the entropy node atlas of the present invention;

[0037] Figure 3 Schematic diagram of the LGCN neural network model structure of the present invention;

[0038] Figure 4 Visualization of model training results. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that the embodiments in this application and the features in the embodiments can be combined with each other arbitrarily without conflict.

[0040] Example 1: Figure 1-4As shown, according to the first aspect of the embodiment of the present invention, a gearbox fault diagnosis method based on mutual information entropy graph is provided, including: step 1: extracting time series nodes from gearbox vibration data under different states to construct time series node sets under each state; preprocessing the time series node sets under each state respectively to obtain the preprocessed time series node sets under each state; step 2: based on the time series dependency of the time series nodes, using the mutual information entropy graph construction method to mine the association relationship of the time series nodes for the preprocessed time series node sets under each state, and constructing the entropy node graph set accordingly; step 3: introducing LSTM to construct a memory graph neural network model, using the entropy node graph set as the input of the memory graph neural network model, mining the spatiotemporal association information of the gearbox vibration data through training and learning, and realizing the training of the memory graph neural network model; step 4: using the trained memory graph neural network model to diagnose the gearbox data to be diagnosed, identifying the gearbox fault state, so as to realize the gearbox fault diagnosis. It should be noted that before the gearbox data to be diagnosed is diagnosed using the trained memory graph neural network model, the entropy node graph set is constructed according to the above steps 1 and 2.

[0041] Furthermore, the step 1 is specifically as follows: according to a preset segmented data length, the gearbox vibration data under different states are segmented to obtain segmented sub-sample data, which are used as time series nodes to construct a time series node set; the time series node set under each state is normalized and then Fourier transformed to obtain the pre-processed time series node set under each state.

[0042] Further, refer to Figure 2 , the step 2 is specifically as follows: according to the preset number of time series nodes, the preprocessed time series node set in each state is divided into nodes according to the time series to obtain multiple sub-time series node sets in each state; for each sub-time series node set in each state, the joint histogram between any two time series nodes and the edge histogram of the corresponding two time series nodes are calculated respectively; the joint histogram between the two time series nodes and the edge histogram of the corresponding two time series nodes are smoothed respectively and the joint probability distribution and the edge probability distribution are calculated; according to the joint probability distribution and the edge probability distribution, the joint entropy and the edge entropy of the corresponding two time series nodes are calculated; according to the joint entropy and the edge entropy of the two time series nodes, the mutual information entropy of the two time series nodes is calculated; the mutual information entropy is used as the weight of the edge of the two time series nodes to construct an entropy node graph set.

[0043] The technical solution of the present invention adopts an "end-to-end" approach to construct an entropy node atlas related to information entropy to obtain training data. This method avoids the deficiency of relying on expert experience for feature extraction in the traditional method; further, on the one hand, the potential dependencies between fault data are accurately captured through mutual information entropy, which can effectively cope with the challenges of building graph structures for complex equipment. On the other hand, the calculation of mutual information entropy is relatively simple and efficient, which can reduce the computational complexity of graph neural networks on large-scale data and improve the training and reasoning efficiency of the model. Most importantly, mutual information entropy can help filter out noise that is not related to fault diagnosis, enhance the robustness and diagnostic accuracy of the model, and make graph neural networks more superior and reliable in actual industrial fault diagnosis.

[0044] Further, refer to Figure 3 The memory graph neural network model includes: an input layer, which is used to take an entropy node graph set as input; a first graph convolution layer, which is used to extract graph features from the input entropy node graph set, and aggregate the feature information of each time series node and its neighboring time series nodes through a graph convolution operation to obtain a first graph feature; a first BatchNorm layer, which is used to perform batch normalization on the first graph feature to eliminate internal covariate shift; a second graph convolution layer, which is used to extract graph features from the first graph feature after batch normalization, and further aggregate wider neighbor features through graph convolution to obtain higher-order information to obtain a second graph feature; a second BatchNorm layer, which is used to perform batch normalization on the second graph feature. The FC layer uses a fully connected layer to reduce the feature dimension and integrate the features based on the second graph features processed by batch normalization to obtain the FC layer graph features; the Dropout layer uses a set drop probability to randomly drop some neurons in the FC layer graph features to obtain the Dropout layer graph features; the LSTM layer uses the Dropout layer graph features as time series inputs and expands the features into time steps, dynamically updates the memory through the input gate, forget gate, and output gate mechanisms, and processes the information of node features changing over time to fuse graph features and time series features; the output layer uses the output of the LSTM layer to map the output of the LSTM layer and output the state category of the gearbox vibration data.

[0045] According to a second aspect of an embodiment of the present invention, a gearbox fault diagnosis device based on a mutual information entropy graph is provided, comprising: an acquisition module for extracting time series nodes from gearbox vibration data under different states to construct a time series node set under each state; preprocessing the time series node set under each state respectively to obtain the preprocessed time series node set under each state; a construction module for mining the association relationship of the time series nodes based on the time series dependency of the time series nodes, using the mutual information entropy graph construction method for the preprocessed time series node set under each state, and constructing an entropy node graph set accordingly; a training module for introducing LSTM to construct a memory graph neural network model, using the entropy node graph set as the input of the memory graph neural network model, mining the spatiotemporal association information of the gearbox vibration data through training and learning, and realizing the training of the memory graph neural network model; a diagnosis module for diagnosing the gearbox data to be diagnosed using the trained memory graph neural network model, identifying the gearbox fault state, so as to realize the gearbox fault diagnosis. The term "module" as used above can be a combination of software and / or hardware to realize the predetermined function. Although the system described in the above embodiment is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceived. For parts of each module that are not described in detail, please refer to the relevant descriptions of other embodiments.

[0046] According to a third aspect of an embodiment of the present invention, there is provided a terminal comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute any one of the above-described methods for digitally simulating the scraping surface morphology.

[0047] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0048] Example 2: Figure 1-4 As shown, the optional embodiments of the present invention are described below in conjunction with the accompanying drawings and experimental data:

[0049] A gearbox fault diagnosis method based on mutual information entropy graph, comprising:

[0050] Step 1: Build a time series node set:

[0051] This embodiment takes the public data set Southeast University gearbox data set as an example to introduce the gearbox fault diagnosis method. The experimental platform of the test bench consists of motor, reduction gearbox, brake and other parts. The gearbox fault types include defect, broken tooth, tooth root crack, tooth surface wear, inner ring fault, outer ring fault, roller fault and inner and outer ring combined fault, and two different working conditions of 20HZ-0V and 30HZ-2V are set. There are 8 fault state signals and 1 healthy state signal for 9 classification tasks.

[0052] The length of segmented data is set to 1024 data points, 1000 sub-sample data can be obtained for the healthy state signal, and 1000 sub-sample data can be obtained for each fault state signal; the segmented sub-sample data are used as time series nodes to construct a time series node set; the time series node sets under each state are normalized respectively, and the normalized data samples are used as the preprocessed time series node set.

[0053] Step 2: Entropy node atlas construction:

[0054] According to the preset number of time series nodes, the preprocessed time series node set in each state is divided into nodes according to the time series to obtain multiple sub-time series node sets in each state; for each sub-time series node set in each state, the joint histogram between any two time series nodes and the corresponding edge histogram of the two time series nodes are calculated respectively; the joint histogram between the two time series nodes and the corresponding edge histogram of the two time series nodes are smoothed respectively and the joint probability distribution and the edge probability distribution are calculated; according to the joint probability distribution and the edge probability distribution, the joint entropy and edge entropy of the corresponding two time series nodes are calculated; according to the joint entropy and edge entropy of the two time series nodes, the mutual information entropy of the two time series nodes is calculated; the mutual information entropy is used as the weight of the edge of the two time series nodes to construct an entropy node graph set, and the mutual information entropy graph construction method is as follows: Figure 2 First, calculate the joint histogram between any two time series nodes under the same sub-time series node set; the calculation formula for the joint histogram of any time series node X and time series node Y is for:

[0055] ;

[0056] in, represents the number of co-occurrences or joint numbers of the i-th discretization interval of the time series node X and the j-th discretization interval of the time series node Y in the discretization space; Used to determine whether the X value and Y value of the timing node in the Kth sub-timing node set fall within the specified discretization interval; and Yes and The index mapped to the discretized space, and is the discretization step size, and Indicates the value of the kth dimension of the timing node X and the timing node Y in the current sub-timing node set;

[0057] Secondly, calculate the edge histogram of two time series nodes X and Y under the same sub-time series node set and :

[0058] ;

[0059] ;

[0060] in, and Represents the separate edge counts used to calculate timing node X and timing node Y, respectively. is the count of the i-th discretization interval of the time series node X in the discretization space, is the count of the jth discretization interval of the time series node Y in the discretization space;

[0061] Then, the joint histogram and edge histogram of the two time series nodes X and Y are smoothed and the joint probability distribution is calculated. and the marginal probability distribution and :

[0062] ;

[0063] ;

[0064] ;

[0065] in, is the joint histogram after smoothing The sum of all counts in , , Respectively represent the edge histogram after smoothing , 'The sum of all counts; It should be noted that the advantages of using smoothed joint histograms and marginal histograms to calculate probability distributions are as follows: First, reducing data sparsity issues: The grid of the joint histogram may contain fewer data points, resulting in discontinuous distribution or too many blank areas; and through smoothing, the errors caused by data sparsity can be reduced, making the joint distribution more coherent. Second, enhancing correlation identification: Noise will affect the observation and analysis of the relationship between variables; and after smoothing, the correlation or related trend between two variables can be better identified. Third, improving the interpretability of the joint distribution: The smoothed joint histogram can more clearly show the joint distribution between the two variables, helping to discover potential patterns and trends.

[0066] After that, calculate the joint entropy of time series node X and time series node Y and marginal entropy , :

[0067] ;

[0068] ;

[0069] ;

[0070] in, represents the joint uncertainty of timing node X and timing node Y, and Represent the individual uncertainties of time series node X and time series node Y respectively;

[0071] Finally, the mutual information entropy between the time series node X and the time series node Y is calculated. The mutual information entropy is:

[0072] ;

[0073] ;

[0074] in, For timing nodes and timing nodes The weight of the edge between them.

[0075] Exemplarily, according to the data given in step one, the dimension of each time series node is 1024, and 10 time series nodes construct a sub-time series node set, that is, ten segments of sub-sample data can construct 1 sub-time series node set. Therefore, each state has a total of 100 sub-time series node sets, and 9 states have a total of 900 sub-time series node sets. Finally, 900 entropy node graphs are constructed as the entropy node graph set.

[0076] Step 3: Memory graph neural network model construction and training:

[0077] The memory graph neural network model constructed by combining LSTM and GCN is as follows Figure 3 As shown (i.e., LGCN neural network model), the LGCN neural network model includes two graph convolution layers (GCNConv), two BatchNorm layers (batch normalization layers), FC layers (fully connected layers), Dropout layers (discard layers), LSTM layers (long short-term memory network layers), and output layers. The specific settings are as follows: the output feature dimension of the first graph convolution layer is 1024; the output feature dimension of the first BatchNorm layer is 1024; the output feature dimension of the second graph convolution layer is 1024; the output feature dimension of the second BatchNorm layer is 1024; the input feature dimension of the FC layer is 1024, and the output feature dimension is 512; the dropout rate of the Dropout layer is set to 0.2; the number of LSTM layers is 1, the input feature dimension is 512, and the output feature dimension is 128; the input feature dimension of the output layer is 128, and the output dimension is the number of fault types, which is set to 9 in this embodiment.

[0078] Among them, the entropy node atlas is divided and labeled. The proportion of labeled training sets is 0.8, the proportion of labeled test sets is 0.2, the initial learning rate is set to 0.001, the parameters are updated by back propagation and Adam optimization algorithm, and the training process adopts mini-batch learning, and the size is set to 32. The loss function uses the cross entropy loss function. While training the model with labeled training sets, the labeled test sets are verified to obtain the best accuracy of the model to test new data that are not involved in training. The labels are as follows: 0 represents the healthy state of the gearbox, 1 represents gearbox defect, 2 represents gearbox broken teeth, 3 represents gearbox tooth root crack, 4 represents gearbox tooth surface wear, 5 represents bearing inner ring fault, 6 represents bearing outer ring fault, 7 represents bearing roller fault, and 8 represents bearing inner ring + outer ring composite fault.

[0079] The network is trained with the gearbox vibration data in the embodiment of the present invention to obtain the best accuracy of the model. The model accuracy calculation method is as follows: the entropy node graph with the true label is used as the memory graph neural network model input, and the predicted label of each entropy node graph is obtained. By comparing the predicted label with the true label, the number of correctly predicted entropy node graphs is counted, and then the number of correctly predicted entropy node graphs is divided by the total number of entropy node graphs to obtain the final accuracy of the model. In addition, the confusion matrix is ​​used to visualize the model output results. The visualization results are as follows: Figure 4As shown in the figure, in the confusion matrix, the horizontal axis represents the predicted labels of each state of the labeled test set, and the vertical axis represents the true labels of each state of the labeled test set. The numerical value represents the proportion of all correctly predicted entropy node graphs to the total number of entropy node graphs. Figure 4 It can be concluded that the model can basically accurately identify the label of each entropy node graph in the test set.

[0080] At the same time, the gearbox vibration data is input into different graph construction methods to obtain the best accuracy of the model and compare it with the method of the present invention to verify the superiority of the present invention. In the embodiment of the present invention, the graph construction methods selected for comparison are: the pre-processed time series node set is directly constructed by the KNN graph construction method under the graph structure feature set (i.e., KNN+GCN), the pre-processed time series node set is directly constructed by the Radius graph construction method under the graph structure feature set (i.e., Radius+GCN), the pre-processed time series node set is directly constructed by the mutual information entropy graph construction method under the graph structure feature set (i.e., mutual information entropy+GCN), and the present invention (i.e., mutual information entropy+LGCN).

[0081] The memory graph neural network fault diagnosis model based on the mutual information entropy graph and the comparison graph construction model were trained 100 times according to the above parameters and the experimental results shown in Table 1 were obtained.

[0082] Table 1 Model experimental results

[0083]

[0084] It can be seen from the experimental results in Table 1 in the embodiments of the present invention that by comparing the experimental results of KNN+GCN, Radius+GCN and mutual information entropy+GCN, it can be seen that the use of the mutual information entropy graph construction method can effectively mine the nonlinear dependencies of time series nodes and improve the effectiveness of feature selection, thereby improving the accuracy of the model; by comparing the model accuracies of mutual information entropy+GCN and mutual information entropy+LGCN, it can be seen that adding an LSTM network to the graph neural network can effectively capture the timing dependencies of timing nodes, enhance information propagation between nodes, and significantly improve the accuracy and reliability of fault diagnosis.

[0085] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A gearbox fault diagnosis method based on mutual information entropy graph, It is characterized in that include: Step 1: Extract time series nodes from the gearbox vibration data under different states to construct a time series node set under each state; Preprocess the time series node sets in each state respectively to obtain the preprocessed time series node sets in each state; Step 2: Based on the temporal dependency of the temporal nodes, the mutual information entropy graph construction method is used to mine the correlation relationship of the temporal nodes for the preprocessed temporal node sets in each state, and an entropy node graph set is constructed accordingly; Step 3: Introduce LSTM to build a memory graph neural network model, use the entropy node graph set as the input of the memory graph neural network model, and mine the spatiotemporal correlation information of the gearbox vibration data through training and learning to realize the training of the memory graph neural network model; Step 4: Use the trained memory graph neural network model to diagnose the gearbox data to be diagnosed, identify the gearbox fault state, and realize the gearbox fault diagnosis; The step 2 is specifically as follows: According to the preset number of time series nodes, the preprocessed time series node set in each state is divided into nodes according to the time series to obtain multiple sub-time series node sets in each state; For each sub-series node set in each state, the joint histogram between any two sequence nodes and the edge histogram of the corresponding two sequence nodes are calculated respectively; The joint histogram between two time series nodes and the corresponding edge histograms of the two time series nodes are smoothed respectively, and the joint probability distribution and the edge probability distribution are calculated; According to the joint probability distribution and the marginal probability distribution, the joint entropy and marginal entropy of the corresponding two time series nodes are calculated; According to the joint entropy and edge entropy of two time series nodes, the mutual information entropy of two time series nodes is calculated; The mutual information entropy is used as the weight of the edge between two time series nodes to construct an entropy node graph. The memory graph neural network model comprises: Input layer, used to take entropy node graph as input; The first graph convolution layer is used to extract graph features from the input entropy node graph set, and aggregate the feature information of each time series node and its neighboring time series nodes through the graph convolution operation to obtain the first graph feature; The first BatchNorm layer is used to perform batch normalization on the first graph features to eliminate internal covariate shift; The second graph convolution layer is used to extract graph features from the first graph features after batch normalization, further aggregate wider neighbor features through graph convolution, and obtain higher-order information to obtain the second graph features; The second BatchNorm layer is used to perform batch normalization on the second graph features; FC layer, based on the batch normalized second graph features, the fully connected layer is used to reduce the feature dimension and integrate the features to obtain the FC layer graph features; The Dropout layer is used to randomly discard some neurons in the FC layer graph features according to the set drop probability to obtain the Dropout layer graph features; LSTM layer, which is used to input the graph features of the Dropout layer as time series and expand the features into time steps. It dynamically updates the memory through the input gate, forget gate, and output gate mechanism, and processes the information of node features changing over time to fuse graph features and time series features. The output layer is used to map the output of the LSTM layer and output the state category of the gearbox vibration data.

2. The gearbox fault diagnosis method based on the mutual information entropy graph according to claim 1, It is characterized in that The step 1 is specifically as follows: according to the preset segmented data length, the gearbox vibration data under different states are segmented to obtain the segmented sub-sample data, which are used as time series nodes to construct a time series node set; the time series node set under each state is normalized and then Fourier transformed to obtain the pre-processed time series node set under each state.

3. A gearbox fault diagnosis device based on mutual information entropy graph, It is characterized in that include: An acquisition module is used to extract time series nodes from the gearbox vibration data under different states to construct a time series node set under each state; Preprocess the time series node sets in each state respectively to obtain the preprocessed time series node sets in each state; A construction module is used to mine the correlation relationship of the time series nodes based on the time series dependency relationship of the time series nodes, use the mutual information entropy graph construction method to build the time series node set preprocessed in each state, and construct the entropy node graph set accordingly; The training module is used to introduce LSTM to build a memory graph neural network model, use the entropy node graph set as the input of the memory graph neural network model, and mine the spatiotemporal correlation information of the gearbox vibration data through training learning to realize the training of the memory graph neural network model; The diagnostic module is used to diagnose the gearbox data to be diagnosed using the trained memory graph neural network model, identify the gearbox fault state, and realize the gearbox fault diagnosis; The building blocks are specifically: According to the preset number of time series nodes, the preprocessed time series node set in each state is divided into nodes according to the time series to obtain multiple sub-time series node sets in each state; For each sub-series node set in each state, the joint histogram between any two sequence nodes and the edge histogram of the corresponding two sequence nodes are calculated respectively; The joint histogram between two time series nodes and the corresponding edge histograms of the two time series nodes are smoothed respectively, and the joint probability distribution and the edge probability distribution are calculated; According to the joint probability distribution and the marginal probability distribution, the joint entropy and marginal entropy of the corresponding two time series nodes are calculated; According to the joint entropy and edge entropy of two time series nodes, the mutual information entropy of two time series nodes is calculated; The mutual information entropy is used as the weight of the edge between two time series nodes to construct an entropy node graph. The memory graph neural network model comprises: Input layer, used to take entropy node graph as input; The first graph convolution layer is used to extract graph features from the input entropy node graph set, and aggregate the feature information of each time series node and its neighboring time series nodes through the graph convolution operation to obtain the first graph feature; The first BatchNorm layer is used to perform batch normalization on the first graph features to eliminate internal covariate shift; The second graph convolution layer is used to extract graph features from the first graph features after batch normalization, further aggregate wider neighbor features through graph convolution, and obtain higher-order information to obtain the second graph features; The second BatchNorm layer is used to perform batch normalization on the second graph features; FC layer, based on the batch normalized second graph features, the fully connected layer is used to reduce the feature dimension and integrate the features to obtain the FC layer graph features; The Dropout layer is used to randomly discard some neurons in the FC layer graph features according to the set drop probability to obtain the Dropout layer graph features; LSTM layer, which is used to input the graph features of the Dropout layer as time series and expand the features into time steps. It dynamically updates the memory through the input gate, forget gate, and output gate mechanism, and processes the information of node features changing over time to fuse graph features and time series features. The output layer is used to map the output of the LSTM layer and output the state category of the gearbox vibration data.

Images