A Fault Diagnosis Method for Insulated Bearings Based on a Residual Space-Time Information Fusion Model

Abstract

A fault diagnosis method for an insulated bearing based on a residual spatio-temporal information fusion model of the present invention includes: using two three-axis vibration sensors to collect the original vibration signals of six channels of the insulated bearing of a doubly-fed wind generator to capture the information of the insulated bearing in different health states; dividing the original vibration signals into 80% training set data and 20% test set data; constructing a graph knowledge empowered residual spatio-temporal information fusion model; optimizing the hyperparameters of the graph knowledge empowered residual spatio-temporal information fusion model by using a scavenger algorithm; taking the training set data as the input of the graph knowledge empowered residual spatio-temporal information fusion model and the fault category as the output to perform model training; inputting the test set data into the trained graph knowledge empowered residual spatio-temporal information fusion model, and through an automatic spatio-temporal feature residual learning process, classifying the fused feature data to achieve intelligent diagnosis of the insulated bearing of the doubly-fed wind generator.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis, and relates to an insulating bearing fault diagnosis method based on a residual space-time information fusion model. Background Art

[0002] Wind turbines are known as the "heart" of wind turbines, and insulated bearings are crucial components supporting the generator's transmission structure. However, high-power variable-frequency generators can cause oil film breakdown, generating shaft current discharges that significantly shorten the bearing fatigue wear process. Statistics show that this failure mode is common in high-power variable-frequency motor bearings. Furthermore, the complex and variable operating conditions of wind turbines make them highly susceptible to failure. Bearing failure can impact the safe operation of the entire wind turbine system and even cause significant economic losses. Therefore, research is needed to diagnose wind turbine bearing faults, monitor and diagnose bearing failures in a timely manner, and formulate inspection and maintenance strategies to ensure the safe and stable operation of wind turbines.

[0003] In recent years, deep learning has been able to fit real-world physical models with powerful nonlinear models, explore nonlinear relationships between variables, and achieve end-to-end fault diagnosis. For example, convolutional neural networks, relying on unique parameter sharing and local connection mechanisms, can achieve important translation invariance in Euclidean space. However, the vast majority of real-world scenarios have distinct irregular topological structures, also known as non-Euclidean data, making traditional neural network models for Euclidean data inadequate. Graph neural networks can collect information from a node's neighbors and use it to update its own implicit state, thereby enabling the propagation of high-order structural information and generally significantly improving the model's expressive power. However, due to the over-correlation between different feature dimensions in traditional graph network models, the problem of high information redundancy and low content of important information in neighbor gain is easily encountered.

[0004] At the same time, given the unclear damage mechanism of the insulated bearings of high-power variable-frequency controlled megawatt-class wind turbines due to the interaction between shaft current and fatigue wear, identifying their electrical corrosion failures has become a pressing technical challenge. Considering that the characteristics of insulated bearing electrical corrosion failures are affected by factors such as mechanical load, bearing structural parameters, and electrical performance parameters, the insulated bearing data is highly complex, nonlinear, and time-varying. Existing models have limited ability to extract temporal information, making it difficult to fully exploit the characteristic information of electrical corrosion failures.

[0005] Moreover, considering that a deep learning model consists of multiple neurons, the parameter space of the model is usually extremely large, containing a large number of weights and biases. In particular, the hyperparameters of deep learning models usually need to be adjusted through manual trial-and-error methods, involving multiple attempts and repeated experiments to adjust the model parameters, resulting in strong randomness and uncertainty, leading to low efficiency in adjusting the hyperparameters of deep learning models, as well as poor recognition accuracy and generalization performance of the models. Summary of the Invention

[0006] To solve the above technical problems, the purpose of the present invention is to provide an insulation bearing fault diagnosis method based on a residual spatio-temporal information fusion model.

[0007] The present invention provides an insulation bearing fault diagnosis method based on a residual spatio-temporal information fusion model, including:

[0008] Step 1: Use 2 triaxial vibration sensors to collect the original vibration signals of 6 channels of the insulation bearing of a doubly-fed wind turbine to capture insulation bearing information in different health states;

[0009] Step 2: Divide the original vibration signals into 80% training set data and 20% test set data;

[0010] Step 3: Construct a graph knowledge empowered residual spatio-temporal information fusion model;

[0011] Step 4: Optimize the hyperparameters of the graph knowledge empowered residual spatio-temporal information fusion model using a scavenger algorithm;

[0012] Step 5: Use the training set data as the input of the graph knowledge empowered residual spatio-temporal information fusion model and the fault category as the output to train the model;

[0013] Step 6: Input the test set data into the trained graph knowledge empowered residual spatio-temporal information fusion model. Through an automatic spatio-temporal feature residual learning process, perform fault classification on the fused feature data to achieve intelligent diagnosis of the insulation bearing of the doubly-fed wind turbine.

[0014] The insulation bearing fault diagnosis method based on a residual spatio-temporal information fusion model of the present invention has at least the following beneficial effects:

[0015] Considering the diagnosis problem of the electric erosion fault of the insulation bearing of a megawatt-level wind turbine with high-power frequency conversion control, the present invention has developed a graph knowledge empowered residual spatio-temporal information fusion model in a unique way, aiming to solve the problem that the electric erosion fault caused by the shaft current damage of the insulation bearing in actual engineering cannot be accurately identified. The main conclusions obtained are as follows:

[0016] (1) Inspired by the structure of biological neurons, a new graph structure was innovatively designed, and the formula for the maximum mutual information coefficient was derived. A new adjacency matrix was constructed and its validity was proven. Based on this graph structure, a spatial information fusion module was designed to calculate the non-Euclidean distance relationship between multi-channel insulating bearing vibration signals, giving the residual space-time information fusion model framework strong reasoning and expression capabilities. This provides a powerful modeling tool for the processing and analysis of unstructured data, and also offers a new perspective for the design of learning algorithms such as graph neural networks.

[0017] (2) Considering the strong time-varying nature of the temporal characteristics of wind turbine insulation bearing data, a bidirectional WDRU neural network is proposed, and the backpropagation formula is updated. This is cleverly applied to the proposed time information fusion module to deeply mine the temporal information of insulation bearings. The features extracted by the spatial information fusion module and the temporal information fusion module are fused to achieve information complementarity.

[0018] (3) In the insulated bearing fault classification stage, the scavenger algorithm is used to optimize the number of hidden layer neurons, learning rate, regularization coefficient and training times of the residual space-time information fusion model enabled by graph knowledge, taking advantage of its fast convergence speed and strong optimization ability. This makes the residual space-time information fusion model enabled by graph knowledge have better classification performance in dealing with the classification problem of time series. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of an insulating bearing fault diagnosis method based on a residual space-time information fusion model of the present invention;

[0020] Figure 2 is a schematic diagram of a weight-reduction recurrent unit;

[0021] Figure 3 It is a schematic diagram of a bidirectional WDRU neural network. DETAILED DESCRIPTION

[0022] like Figure 1 As shown, the present invention provides an insulating bearing fault diagnosis method based on a residual space-time information fusion model, comprising:

[0023] Step 1: Use two three-axis vibration sensors to collect six channels of raw vibration signals from the insulating bearing of the doubly-fed wind turbine to capture information about the insulating bearing in different health states.

[0024] Step 2: Divide the original vibration signal into 80% training set data and 20% test set data.

[0025] Step 3: Build a residual spatiotemporal information fusion model empowered by graph knowledge.

[0026] The residual spatio-temporal information fusion model empowered by graph knowledge includes: a spatial information fusion module, a temporal information fusion module, and a self-attention module; the original vibration signals of 6 channels are respectively input into the spatial information fusion module and the temporal information fusion module to extract local features and global features, and the local features and global features are fused in the self-attention module to obtain fused features, and the fused features are input into a fully connected layer to construct a mapping relationship between features and fault modes and output a diagnosis result.

[0027] In specific implementation, the spatial information fusion module includes: a first hidden layer and a second hidden layer. The first hidden layer includes: a first graph convolutional layer and a first activation layer, and the first graph convolutional layer includes 64 graph convolutional kernels; the second hidden layer includes: a second graph convolutional layer, a batch normalization layer, a second activation layer, and a dropout layer, and the second graph convolutional layer includes 128 graph convolutional kernels; each graph convolutional kernel performs a weighted sum on a central vertex and its surrounding 23 adjacent nodes; the original vibration signals of 6 channels are sequentially input into the first graph convolutional layer, the first activation layer, the second graph convolutional layer, the batch normalization layer, the second activation layer, and the dropout layer, and local features are output.

[0028] The graph convolutional network can merge spatial information through graph convolution. In particular, graph neural networks show excellent capabilities in modeling the feature and structural information of graphs. However, there is a new phenomenon when graph neural networks capture more information from high-order neighbors, which is called Overcrrelation. Overcrrelation will make the captured representation contain a large amount of redundant information and cause the loss of some useful information, thereby reducing the expressive ability of the model. Therefore, in order to solve the above technical challenges existing in the graph network model, and at the same time to fully mine the wind turbine insulation bearing data with non-Euclidean spatial relationships generated by triaxial vibration sensors. The present invention introduces a graph convolutional network based on the spatial domain to describe the non-Euclidean spatial relationship between 6 directions of 2 triaxial vibration sensors. First, the Euclidean data is converted into graph data considering the neighbor relationship, and the spatial information fusion module is obtained through graph convolution. The operation process of this module is to learn the interaction between attribute features by performing an affine transformation on the attribute information, and then iteratively aggregate the features of neighbor vertices to update the features of the current vertex. Each time a graph convolution operation is added, the central vertex can fuse more information from external vertices.

[0029] Calculating the correlation matrix is a prerequisite for constructing the adjacency matrix. The adjacency matrix is used to store the information in the set of edges between nodes in graph data and is a two-dimensional array. In the present invention, constructing the adjacency matrices of the graph convolutional layers in the first hidden layer and the second hidden layer, and obtaining the output of the graph convolutional kernels specifically includes the following steps:

[0030] (1) Calculate the maximum mutual information coefficient of the original vibration signal sequences of adjacent channels:

[0031]

[0032]

[0033]

[0034]

[0035] Among them, X and Y respectively represent the original vibration signal sequences of two channels, M IC (X, Y) is the maximum mutual information coefficient between the X sequence and the Y sequence, I(X; Y) represents the mutual information between the X sequence and the Y sequence, H(X) represents the entropy of the X sequence, and H(Y) represents the entropy of the Y sequence; p(x, y) represents the joint probability of the elements in the X sequence and the Y sequence, p(x) represents the probability of a certain element appearing in the X sequence, and p(y) represents the probability of a certain element appearing in the Y sequence; p(x i ) represents the probability of the i-th element appearing in the X sequence, p(y i ) represents the probability of the i-th element appearing in the Y sequence, I(x i ) represents the entropy of the i-th element in the X sequence, I(y i ) represents the entropy of the i-th element in the Y sequence, and n represents the number of elements included in the sequence;

[0036] In order to better quantify the value range of the mutual information coefficient between the original vibration signal sequences of any two channels, when the original vibration signal sequences of the two channels are the same, it is defined that the maximum self-information between the two channels is 1, so that in the adjacency matrix, the relationship range between the two channels is [0, 1].

[0037] When the X sequence and the Y sequence are the same, it is a self-information relationship. Now the proof is as follows:

[0038] p(x, y) = p(x) = p(y)

[0039]

[0040] Therefore:

[0041]

[0042] (2) Use the maximum mutual information coefficient to construct the adjacency matrix G of the graph convolutional layer in the first hidden layer and the second hidden layer as follows:

[0043]

[0044] Among them, Ax, Ay, Az, Bx, By, and Bz respectively represent the original vibration signal sequences of 6 channels;

[0045] (3) Calculate the normalized adjacency matrix using Laplace

[0046]

[0047]

[0048] where \(E\) is the identity matrix, is the degree matrix of;

[0049] (4) Obtain the output of the graph convolution kernel of the \((l + 1)\)-th layer:

[0050]

[0051] In the formula: \(H\) l is the output of the graph convolution kernel of the \(l\)-th layer; \(f(\cdot)\) is the non-linear activation function; \(W\) l+1 is the trainable weight matrix of the \((l + 1)\)-th layer.

[0052] The gated recurrent unit is a variant of the recurrent neural network, which can better capture the dependencies at larger time step distances in the time series. Inspired by the improved idea of the gated recurrent unit, the weights of the LSTM model are uniformly controlled by a gate structure weight, and the weight-reduced recurrent unit (WDRU) is proposed, which can greatly reduce the number of weights to be optimized in the neural network, thereby ensuring that the training time of the model on the data is reduced without reducing the classification accuracy of the model.

[0053] Specifically, when implemented, the time information fusion module includes: an input layer, a third hidden layer, and an output layer. As Figure 2 shown, the third hidden layer includes a bidirectional WDRU neural network. Through the bidirectional WDRU neural network, forward and backward propagation are performed to extract global features, which can better capture the bidirectional dependencies in the time series data. The forward WDRU processes the input sequence from start to end, while the backward WDRU processes it in the opposite direction. The WDRU neural network includes multiple weight-reduced recurrent units, and the schematic diagram of the weight-reduced recurrent unit is as Figure 3 shown. Its main design idea is to reduce the gate structure to a shared gate, thereby effectively streamlining the framework structure and improving the model training speed.

[0054] The calculation process of the bidirectional WDRU neural network is as follows:

[0055] (1) Calculate the information state of the shared gate:

[0056] f t = W d · [h t-1 , xt +b d

[0057] D t =σ(f t )=σ(W d ·[h t-1 ,x t +b d )=σ(W d ·h t-1 +W d ·x t +b d )

[0058] Q t =tan(f t )=tan(W d ·[h t-1 ,x t +b d )=tan(W d ·h t-1 +W d ·x t +b d )

[0059] where f t represents the intermediate variable of the weight reduction recurrent unit, W d is the weight parameter of the weight reduction recurrent unit, h t-1 represents the output of the previous weight reduction recurrent unit, x t represents the original vibration signal of a certain channel, b d represents the bias parameter of the weight reduction recurrent unit; D t represents the shared gate, Q t represents the information state of the shared gate;

[0060] (2) Update the control parameter of the information state of the weight reduction recurrent unit: C t =D t ×C t-1 +(1 - D t )×Q t =σ(W d ·[h t-1 ,x t +b d )×C t-1 +(1 - σ(f t ))×tan(f t )

[0061] where C t-1 represents the control parameter of the information state of the previous weight reduction recurrent unit; C tA control parameter indicating the information state of the current weight reduction recurrent unit;

[0062] (3) Calculate the output of the weight reduction recurrent unit:

[0063] h t = D t × tan(C t ) = σ(W d · h t-1 + W d · x t + b d ) × tan(C t )

[0064] (4) Calculate the output of the bidirectional WDRU neural network:

[0065] y t = σ(h t ) = σ(D t × tan(C t ))

[0066] (5) The update formula of the bidirectional WDRU neural network is as follows:

[0067] h t = W Ft × h Ft + W Bt × h Bt + b t

[0068] where, W Ft and W Bt respectively represent the weight variables in the forward and backward propagation processes, b t represents the bias, h Ft represents the forward information state, h Bt represents the backward information state,

[0069] The backward propagation formula of the bidirectional WDRU neural network is as follows:

[0070] (6) The minimum optimization objective of the time information fusion module is:

[0071]

[0072] where, E t represents the error value at the t-th time period, Y t represents the output of the time information fusion module; W y represents the weight parameter of the weight reduction recurrent unit; b y represents the bias parameter of the weight reduction recurrent unit;

[0073] (7) Variables of the output layer of the time information fusion module:

[0074]

[0075]

[0076]

[0077]

[0078] Among them, P t represents the output of the current time weight reduction recurrent unit; δy t represents the change in the output of the bidirectional WDRU neural network model; δP t The change in the output of the current time weight reduction recurrent unit; δW y represents the change in the weight parameters of the bidirectional WDRU neural network; δb y represents the change in the bias parameter of the bidirectional WDRU neural network;

[0079] (8) Calculate the change in weights during the backpropagation process of the hidden layer of the time information fusion module:

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] Among them, δh t δC t δD t δQ t δf t w h all represent the change in weights during the backpropagation process of the network model, and T represents the period of the time series signal.

[0086] The Dropout mechanism is added to the hidden layer of the time information fusion module for module optimization, so that neurons have a certain probability of inactivation, thereby making the module have stronger generalization ability and avoiding overfitting of the module. At the same time, the ReLU function is selected as the activation function in this module, so that the constructed residual spatio-temporal information fusion model empowered by graph knowledge has strong anti-interference ability and regularity, and at the same time avoids the problem of gradient disappearance.

[0087] Step 4: Optimize the hyperparameters of the residual spatio-temporal information fusion model empowered by graph knowledge using the scavenger algorithm, thereby improving the recognition accuracy and generalization ability of the model. The hyperparameters include: the number of neurons in the first hidden layer, the second hidden layer, and the third hidden layer of the residual spatio-temporal information fusion model empowered by graph knowledge, the learning rate, the regularization coefficient, the number of iterations, and the batch size.

[0088] Specifically, when implementing, optimize the hyperparameters of the residual spatio-temporal information fusion model empowered by graph knowledge using the scavenger algorithm, specifically as follows:

[0089] (1) Initialize the residual spatio-temporal information fusion model empowered by graph knowledge, including setting the number of neurons in the hidden layer of the initial residual spatio-temporal information fusion model empowered by graph knowledge.

[0090] (2) Set the ranges of the learning rate, the number of iterations, and the batch size of the residual spatio-temporal information fusion model empowered by graph knowledge. It is also necessary to set the scavenger target tracking speed coefficient, the scavenger population size, and the maximum number of iterations.

[0091] (3) Calculate the fitness values of all scavenger individuals, sort them according to the fitness size, and select the individual with the best fitness. Select the mean square error as the fitness function of the scavenger algorithm. MSE is expressed as:

[0092]

[0093] where n is the number of samples, Y i and Y i * are the true value and the predicted value of the i-th sample data, respectively.

[0094] The smaller the MSE value, the smaller the gap between the predicted value and the true value of the residual spatio-temporal information fusion model empowered by graph knowledge, indicating that the position of the scavenger individual is better; the position of the scavenger individual corresponds one-to-one with the hyperparameters of the residual spatio-temporal information fusion model empowered by graph knowledge. The position coordinates of the scavenger individual with the best fitness are the optimal hyperparameters of the residual spatio-temporal information fusion model empowered by graph knowledge.

[0095] (4) Determine whether the scavenger algorithm has iterated to the end. When the maximum number of iterations is reached, the algorithm iteration is completed. At this time, assign the position of the scavenger individual with the best fitness to the residual spatio-temporal information fusion model empowered by graph knowledge, that is, assign the optimal hyperparameters found by the scavenger algorithm to the residual spatio-temporal information fusion model empowered by graph knowledge. If the scavenger algorithm iteration has not ended, return to the previous step.

[0096] Step 5: Use the training set data as the input of the residual spatio-temporal information fusion model empowered by graph knowledge, and the fault category as the output to train the model.

[0097] In specific implementation, the training set data is used as the input of the residual spatio-temporal information fusion model empowered by graph knowledge, and the fault category is used as the output to complete the feature extraction and fusion of the original vibration data of the insulated bearing. The loss function of the Adam optimizer and the Softmax classifier is selected, and the training is stopped after reaching the set number of iterations, and the fault type is output to complete the model training.

[0098] Step 6: Input the test set data into the trained residual spatio-temporal information fusion model empowered by graph knowledge. Through the automatic spatio-temporal feature residual learning process, the fused feature data is fault classified to realize the intelligent diagnosis of the insulated bearing of the doubly-fed wind turbine.

[0099] The above are only the preferred embodiments of the present invention and are not intended to limit the idea of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A fault diagnosis method for insulated bearings based on a residual space-time information fusion model, characterized in that Including: Step 1: Use two triaxial vibration sensors to collect the original vibration signals of 6 channels of the insulated bearings of a doubly-fed wind turbine to capture the information of insulated bearings in different health states; Step 2: Divide the original vibration signals into 80% training set data and 20% test set data; Step 3: Construct a graph knowledge empowered residual spatio-temporal information fusion model; Step 4: Use the scavenger algorithm to optimize the hyperparameters of the graph knowledge empowered residual spatio-temporal information fusion model; Step 5: Use the training set data as the input of the graph knowledge empowered residual spatio-temporal information fusion model and the fault category as the output to train the model; Step 6: Input the test set data into the trained graph knowledge empowered residual spatio-temporal information fusion model. After an automatic spatio-temporal feature residual learning process, perform fault classification on the fused feature data to achieve intelligent diagnosis of the insulated bearings of the doubly-fed wind turbine; The graph knowledge empowered residual spatio-temporal information fusion model includes: a spatial information fusion module, a temporal information fusion module, and a self-attention module; the original vibration signals of 6 channels are respectively input into the spatial information fusion module and the temporal information fusion module to extract local features and global features. The local features and global features are fused in the self-attention module to obtain fused features. The fused features are input into a fully connected layer to construct the mapping relationship between features and fault patterns and output the diagnosis result; The spatial information fusion module includes: a first hidden layer and a second hidden layer; the first hidden layer includes: a first graph convolutional layer and a first activation layer, and the first graph convolutional layer includes 64 graph convolutional kernels; the second hidden layer includes: a second graph convolutional layer, a batch normalization layer, a second activation layer, and a dropout layer, and the second graph convolutional layer includes 128 graph convolutional kernels; each graph convolutional kernel performs a weighted sum on a central vertex and its 23 adjacent nodes around it; the original vibration signals of 6 channels are sequentially input into the first graph convolutional layer, the first activation layer, the second graph convolutional layer, the batch normalization layer, the second activation layer, and the dropout layer to output local features; among them, the maximum mutual information coefficient is used to construct the adjacency matrix of the graph convolutional layers in the first hidden layer and the second hidden layer; The temporal information fusion module includes: an input layer, a third hidden layer, and an output layer. The third hidden layer includes a bidirectional WDRU neural network. Through the bidirectional WDRU neural network, forward and backward propagation are performed to extract global features; the WDRU neural network includes multiple weight reduction recurrent units; The calculation process of the bidirectional WDRU neural network is as follows: (1) Calculate the information state of the shared gate: f t = W d · [h t-1 , x t + b d D t = σ(f t ) = σ(W d · [h t-1 , x t + b d ) = σ(W d · h t-1 + W d · x t + b d ) Q t = tan(f t ) = tan(W d · [h t-1 , x t + b d ) = tan(W d · h t-1 + W d · x t + b d ) Among them, f t represents the intermediate variable of the weight reduction recurrent unit, and W d is the weight parameter of the weight reduction recurrent unit, h t-1 represents the output of the previous weight reduction recurrent unit, and x t represents the original vibration signal of a certain channel, and b d represents the bias parameter of the weight reduction recurrent unit; D t represents the shared gate, and Q t represents the information state of the shared gate; (2) Update the control parameters of the information state of the weight reduction recurrent unit: C t = D t × C t-1 + (1 - D t ) × Q t = σ(W d · [h t-1 , x t + b d ) × C t-1 + (1 - σ(f t )) × tan(f t ) where C t-1 represents the control parameter of the information state of the previous weight reduction recurrent unit; C t represents the control parameter of the information state of the current weight reduction recurrent unit; (3) Calculate the output of the weight reduction recurrent unit: h t = D t × tan(C t ) = σ(W d · h t-1 + W d · x t + b d ) × tan(C t ) (4) Calculate the output of the bidirectional WDRU neural network: y t = σ(h t ) = σ(D t × tan(C t )) (5) The update formula of the bidirectional WDRU neural network is as follows: h t = W Ft × h Ft + W Bt × h Bt + b t Among them, W Ft and W Bt represent the weight variables in the forward and backward propagation processes respectively, b t represents the bias, h Ft represents the forward information state, h Bt represents the backward information state.

2. The insulation bearing fault diagnosis method based on the residual space-time information fusion model according to claim 1, characterized in that Construct the adjacency matrix of the graph convolutional layers in the first hidden layer and the second hidden layer, and obtain the output of the graph convolutional kernels, which specifically includes the following steps: (1) Calculate the maximum mutual information coefficient of the original vibration signal sequences of adjacent channels: Among them, X and Y respectively represent the original vibration signal sequences of two channels, M IC (X, Y) is the maximum mutual information coefficient between the X sequence and the Y sequence, I(X; Y) represents the mutual information between the X sequence and the Y sequence, H(X) represents the entropy of the X sequence, and H(Y) represents the entropy of the Y sequence; p(x, y) represents the joint probability of the elements in the X sequence and the Y sequence, p(x) represents the probability of a certain element appearing in the X sequence, and p(y) represents the probability of a certain element appearing in the Y sequence; p(x i ) represents the probability of the i-th element appearing in the X sequence, p(y i ) represents the probability of the i-th element appearing in the Y sequence, I(x i ) represents the entropy of the i-th element in the X sequence, I(y i ) represents the entropy of the i-th element in the Y sequence, and n represents the number of elements included in the sequence; (2)Construct the adjacency matrix \(G\) of the graph convolutional layer in the first hidden layer and the second hidden layer using the maximum mutual information coefficient as follows: where \(A_x\), \(A_y\), \(A_z\), \(B_x\), \(B_y\), and \(B_z\) respectively represent the original vibration signal sequences of 6 channels; (3) Calculate the normalized adjacency matrix using Laplace where \(E\) is the identity matrix, is the degree matrix of; (4)Obtain the output of the \((l + 1)\)-th layer graph convolutional kernel: Where: H l is the output of the l-th layer graph convolution kernel; f(·) is a non-linear activation function; W l+1 is the trainable weight matrix of the (l + 1)-th layer.

3. The insulation bearing fault diagnosis method based on the residual spatio-temporal information fusion model according to claim 1, characterized in that: The backpropagation formula of the bidirectional WDRU neural network is as follows: (6)The minimum optimization objective of the time information fusion module is: Among them, E t represents the error value at the t-th time period, and Y t represents the output of the time information fusion module; W y represents the weight parameter of the weight reduction recurrent unit; b y represents the bias parameter of the weight reduction recurrent unit; (7)The variables of the output layer of the time information fusion module: Among them, P t represents the output of a current time-weight reduction recurrent unit; δy t represents the change in the output of the bidirectional WDRU neural network model; δP t is the change in the output of a current time-weight reduction recurrent unit; δW y represents the change in the weight parameters of the bidirectional WDRU neural network; δb y represents the change in the bias parameters of the bidirectional WDRU neural network; (8)Calculate the weight change amount during the backpropagation of the hidden layer of the time information fusion module: Among them, δh t δC t δD t δQ t δf t w h All represent the change amount of the weights during the backpropagation process of the network model, and T represents the period of the time series signal.

4. The insulation bearing fault diagnosis method based on the residual space-time information fusion model according to claim 1, wherein, In step 4, the scavenger algorithm is used to optimize the hyperparameters of the graph knowledge empowered residual spatio-temporal information fusion model, thereby improving the recognition accuracy and generalization ability of the model. The hyperparameters include: the number of neurons in the first hidden layer, the second hidden layer, and the third hidden layer of the graph knowledge empowered residual spatio-temporal information fusion model, learning rate, regularization coefficient, number of iterations, and batch size.

5. The insulation bearing fault diagnosis method based on the residual space-time information fusion model according to claim 4, characterized in that In step 4, the scavenger algorithm is used to optimize the hyperparameters of the graph knowledge empowered residual spatio-temporal information fusion model, specifically: (1)Initialize the graph knowledge empowered residual spatio-temporal information fusion model, including setting the number of neurons in the initial hidden layer of the graph knowledge empowered residual spatio-temporal information fusion model; (2)Set the range of the learning rate, number of iterations, and batch size of the graph knowledge empowered residual spatio-temporal information fusion model. It is also necessary to set the scavenger target tracking speed coefficient, scavenger population size, and maximum number of iterations; (3)Calculate the fitness values of all scavenger individuals, sort them according to the fitness size, and select the individual with the best fitness. Select the mean square error as the fitness function of the scavenger algorithm. MSE is expressed as: where n is the number of samples, and Y i and are the true value and the predicted value of the i-th sample data, respectively; The smaller the MSE value, the smaller the gap between the predicted value and the true value of the graph knowledge empowered residual spatio-temporal information fusion model, indicating that the position of the scavenger individual is better; the position of the scavenger individual corresponds one-to-one with the hyperparameters of the graph knowledge empowered residual spatio-temporal information fusion model. The position coordinates of the scavenger individual with the best fitness are the best hyperparameters of the graph knowledge empowered residual spatio-temporal information fusion model; (4)Judge whether the scavenger algorithm has finished iterating. When the maximum number of iterations is reached, the algorithm iteration is completed. At this time, the position of the scavenger individual with the best fitness is assigned to the graph knowledge empowered residual spatio-temporal information fusion model, that is, the best hyperparameters found by the scavenger algorithm are assigned to the graph knowledge empowered residual spatio-temporal information fusion model. If the scavenger algorithm iteration has not ended, return to the previous step.

6. The insulation bearing fault diagnosis method based on the residual space-time information fusion model according to claim 1, wherein Step 5 is specifically: using the training set data as the input of the graph knowledge empowered residual spatio-temporal information fusion model and the fault category as the output, complete the feature extraction and fusion of the original vibration data of the insulation bearing, select the loss function of the Adam optimizer and the Softmax classifier, stop training when the set number of iterations is reached, output the fault type, and complete the model training.

Images