Electronic load circuit fault diagnosis method based on multi-channel graph attention network

Abstract

The present application proposes an electronic load circuit fault diagnosis method based on a multi-channel graph attention network. Each vibration sensor signal is evenly divided into several time slices, and the frequency domain information of each slice signal is used as a graph node. The edges are connected according to the feature similarity weight coefficients between the graph nodes. Each vibration sensor signal constitutes a subgraph, thereby establishing a multi-subgraph structure; the graph attention mechanism network is used to aggregate and update the information of each subgraph node; the updated nodes of the same time slice from different subgraphs are spliced and fused in a head-to-tail manner to form a fusion graph, and the edge weight learning mechanism and the multi-channel attention mechanism are used to train the edge connection weights between the fusion graph nodes; the attention mechanism aggregation features of the fusion graph are input into the meta-learning model to perform electronic load circuit abnormality diagnosis. The present application proposes a fault diagnosis method based on a multi-channel graph attention network to realize fault detection of components in electronic load circuits.

Description

Technical Field

[0001] This application belongs to the field of deep learning applications, and specifically designs an electronic load circuit fault diagnosis method based on a multi-channel graph attention network. Background Art

[0002] Electronic loads play a huge role in industrial production. However, if they work under extreme conditions for a long time, various types of faults will occur. In order to improve the robustness of electronic loads and reduce maintenance costs, it is necessary to develop a fault diagnosis method for electronic load circuits.

[0003] Most traditional intelligent fault diagnosis methods only consider a single sensor. This approach is widely used in small devices, but impractical for large ones. Traditional fault diagnosis models such as CNNs, RNNs, and LSTMs fail to exploit correlations between signals. While graph-based methods like GNNs and GCNs can effectively mine signal features, relationships between signals, and topological information, they ignore the importance differences between different nodes in the graph. This paper proposes a multi-channel graph attention network-based fault diagnosis method for electronic load circuits. By constructing multiple subgraphs from multiple sensors and employing a multi-head graph attention mechanism to aggregate and update node features, this method effectively enhances the ability to distinguish fault types. Summary of the Invention

[0004] The purpose of this application is to provide an intelligent method to detect the vibration information of electronic loads in large machinery and ensure timely detection of electronic load failures.

[0005] To solve the above technical problems, this application proposes an electronic load circuit fault diagnosis method based on a multi-channel graph attention network, including:

[0006] Each vibration sensor signal is evenly divided into several time slices. The frequency domain information of each slice signal is used as a graph node. Edges are connected between graph nodes based on the feature similarity weight coefficients. Each vibration sensor signal constitutes a subgraph, thus establishing a multi-subgraph structure.

[0007] A multi-head graph attention mechanism is used to aggregate and update the node feature vectors in each subgraph;

[0008] The updated nodes from the same time slice of different subgraphs are spliced and fused in an end-to-end manner to form a fusion graph. The edge connection weights between the nodes of the fusion graph are trained based on the edge weight learning mechanism and the multi-channel attention mechanism.

[0009] The attention mechanism aggregated features of the fusion graph are input into the MAML model, and the fault diagnosis of the electronic load circuit is realized based on the meta-learning model.

[0010] Optionally, each vibration sensor signal is evenly divided into several time slices, and the frequency domain information of each slice signal is used as a graph node. Edges are connected between the graph nodes based on the feature similarity weight coefficients. Each vibration sensor signal constitutes a subgraph, thereby establishing a multi-subgraph structure, including:

[0011] The sensor signal is cut into several time slices so that for each time Sample slice below For example, , is the sample length, is the number of sensors, and the original sample signals from the same sensor are After normalization preprocessing and FFT transformation, it is used as the feature representation of node embedding. In addition, in order to facilitate the subsequent fusion process, the data from different sensors at the same time are uniformly labeled, such as subgraph1.1, subgraph2.1, and subgraph3.1, which represent the data from subgraphs 1 and 2, respectively. Figure 1 , 2, 3 nodes from the first time slice, and then give each node in the same subgraph a learnable parameter (The distance between each two nodes Generally different), according to the parameters after learning Choose whether to connect the edges That is, the two nodes are connected, otherwise they are not connected.

[0012] Optionally, a multi-head graph attention mechanism is used to aggregate and update the node feature vectors in each subgraph, including:

[0013] For each node in the subgraph , you need to first calculate the central nodes two by two The similarity coefficient with its neighbor nodes, and Similarity coefficient between It is defined as, , and Represents nodes respectively and nodes Node characteristics, Represents the feature concatenation operation, matrix and mapping is the parameter to be learned, and then the obtained And the activation function ,function Calculating the attention coefficient , , for The collection of neighbor nodes, and finally the attention coefficient is used to focus on the central node Aggregate with its neighboring nodes, and the node features after aggregation , for Activation function, in order to consider the influence of adjacent nodes on the central node in different aspects, a multi-head graph attention mechanism is used to independently calculate multiple attention heads. The node features calculated by all attention heads are connected to obtain the final aggregated node features. , superscript Indicates the Note the parameters of the head.

[0014] Optionally, the updated nodes from the same time slice of different subgraphs are spliced and fused in an end-to-end manner to form a fusion graph. The edge connection weights between the nodes of the fusion graph are trained based on the edge weight learning mechanism and the multi-channel attention mechanism, including:

[0015] Connect the nodes with the same number end to end and merge them to form a fusion node , ,(c=1,2,3,…,k), then the edge connection between the fusion nodes is completed through the multi-channel attention mechanism, and the average weight of the edge between the nodes of each subgraph is calculated. Form a vector by repeating a scalar value , Number the sub-images as an independent "channel", and then calculate the attention weight of each channel , , , , , is the transformation matrix, is the length of the average matrix, and finally the weight value between the fusion nodes is obtained by weighted summation , are learnable parameters, is the original edge weight between two nodes in each subgraph, according to Choose whether to connect the edges or not. That is, the two nodes are connected, otherwise they are not connected.

[0016] Optionally, the attention mechanism aggregated features of the fusion graph are fed into the MAML model to quickly adapt the meta-tasks for electronic load circuit fault diagnosis, including:

[0017] Using classifiers To perform fault diagnosis operations, , is a meta-parameter, which is trained by inner and outer loops: First, the training task Support set To update the two parameters, , , This is the inner loop, then use the query set To evaluate the updated parameters, the optimization goal of meta-learning is , this is the outer loop, then for the new fault diagnosis task, use the support set To fine-tune the meta parameters , get the updated meta parameters , , and then the total feature matrix of the fusion graph is obtained by passing it through a layer of multi-head graph attention network Input the fully connected layer of the classifier and get , and finally use Calculate the probability of failure, ,The fault type corresponding to the maximum probability is the diagnosis result.

[0018] The present application proposes an electronic load circuit fault diagnosis method based on a multi-channel graph attention network. Each vibration sensor signal is evenly divided into several time slices, and the frequency domain information of each slice signal is used as a graph node. The edges are connected according to the feature similarity weight coefficients between the graph nodes. Each vibration sensor signal constitutes a subgraph, thereby establishing a multi-subgraph structure; the graph attention mechanism network is used to aggregate and update the information of each subgraph node; the updated nodes of the same time slice from different subgraphs are spliced and fused in a head-to-tail manner to form a fusion graph, and the edge weight learning mechanism and the multi-channel attention mechanism are used to train the edge connection weights between the fusion graph nodes; the attention mechanism aggregation features of the fusion graph are input into the meta-learning model to perform electronic load circuit abnormality diagnosis. The present application proposes a fault diagnosis method based on a multi-channel graph attention network to realize fault detection of components in electronic load circuits. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 A schematic diagram of a process for detecting abnormal vibration signals of various sensors in an electronic load circuit is provided for an electronic load circuit fault diagnosis method based on a multi-channel graph attention network according to an embodiment of the present application;

[0021] Figure 2 An electronic load circuit diagram provided by an embodiment of the present invention;

[0022] Figure 3 A diagram showing the working principle of the graph attention network provided for the example of the present invention;

[0023] Figure 4 A diagram showing the working principle of the multi-head graph attention network provided by the example of the present invention. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0025] like Figure 1 Show, Figure 1 The fault diagnosis method for the electronic load circuit based on the multi-channel graph attention network provided in the embodiment of the present application realizes a schematic diagram of the abnormal vibration signal detection process of each sensor of the electronic load circuit, which specifically includes four contents.

[0026] S11: Each vibration sensor signal is evenly divided into several time slices, and the frequency domain information of each slice signal is used as a graph node. Edge connections are made between the graph nodes according to the feature similarity weight coefficients. Each vibration sensor signal constitutes a subgraph, thereby establishing a multi-subgraph structure;

[0027] It should be noted that when a fault occurs, the vibration information collected by multiple sensors in the electronic load circuit is not only related in time, but the information collected by different sensors at the same time is also closely related. If a part of the circuit fails, the information collected by sensors in other parts at the same time will also be affected. This application converts the vibration data set collected by multiple sensors into a multi-subgraph structure for processing.

[0028] Step 11: For Figure 2 Raw data set from sensors in electronic load circuit , slice it evenly in time order to get the original data set under each time slice :

[0029]

[0030] in, is the sample length, is the number of sensors;

[0031] Step 12: Use Normalization preprocesses the raw signals from the same sensor and uses FFT transformation to transform the samples from the time domain to the frequency domain, using the frequency domain information as the feature representation of the node embedding:

[0032]

[0033] Step 13: To facilitate the subsequent fusion process, the data from different sensors at the same time are uniformly labeled, such as subgraph1.1, subgraph2.1, and subgraph3.1, which represent the data from subgraphs 1 and 2. Figure 1 , 2, 3 are nodes from the first time slice;

[0034] Step 14: Assign a learnable parameter to each node in the same subgraph (The distance between each two nodes Generally different), according to the parameter after learning Choose whether to connect the edges That is, the two nodes are connected, otherwise they are not connected.

[0035] Based on the above discussion, in an optional embodiment of the present application, the number of sensors is 3, and the corresponding number of subgraphs obtained is also 3. There are 5 nodes in each subgraph, corresponding to a vibration information sample divided into 5 slices according to time, and the dimension n of each node feature vector is 1000 dimensions.

[0036] S12: Use the multi-head graph attention mechanism to aggregate and update the node feature vectors in each subgraph.

[0037] It should be noted that the nodes in the graph do not exist in isolation, but are interconnected through edges. Aggregating neighbor features can explicitly model the interactions between nodes, so that the representation of the node contains local and even global structural information. The attention mechanism autonomously learns the importance differences of different parts of the input data, so that the model can focus on key information under the guidance of the attention coefficient, and automatically filter out secondary information, thereby achieving a more accurate and efficient feature learning process. Therefore, this application adopts a multi-head graph attention mechanism to aggregate and update the feature vectors of each subgraph node.

[0038] Step 21: Figure 3 As shown, for each subgraph, the central node is The eigenvector of With neighbors Node characteristics of nodes After linear transformation, the concatenation is input into a single-layer neural network to calculate the similarity coefficient. , then use the activation function The similarity coefficient is activated nonlinearly, and finally The function normalizes the similarity after activation and converts the raw score into a probability distribution :

[0039]

[0040] Among them, the coefficient matrix and mapping are the parameter matrices to be learned, Central node The set of neighbor nodes of

[0041] Step 22: The central node uses these attention coefficients The features of neighboring nodes are weighted and summed while retaining their own feature information. Finally, nonlinearity is introduced through the activation function to obtain the new feature representation of the central node:

[0042] in, is the activation function ;

[0043] Step 23: The above steps are the working process of an attention head. In order to enhance the model's ability to express complex graph structures by independently modeling the node relationships in different feature subspaces, and at the same time utilize the redundant complementarity of multi-head information to improve the robustness of the model, such as Figure 4 ,Here we use a multi-head graph attention mechanism that independently calculates multiple attention heads, and obtains the final aggregated node features by connecting the node features calculated by all attention heads:

[0044]

[0045] Among them, the superscript Indicates the Note the parameters of the head.

[0046] Based on the above discussion, in an optional embodiment of the present application, the parameter matrix The dimension is 700*1000, and each subgraph independently calculates 16 attention heads. The feature vector of each node The dimension is 11200.

[0047] S13: The node features of the same time slice of different subgraphs are spliced and fused by connecting them end to end to form a fusion graph. Then, the edge weight learning mechanism is combined with the multi-channel attention mechanism to optimize the edge connection relationship between the nodes in the fusion graph.

[0048] It should be noted that if a part of the electronic load circuit fails, it will definitely affect the vibration signals collected by sensors in other parts. This application adopts the method of fusing multi-channel sub-graphs to fuse the vibration information collected by each sensor at the same time.

[0049] Step 31: Connect the node sequences with the same number end to end and fuse them to generate a fused node:

[0050]

[0051] Where c=1,2,3,…,k, k is the number of sensors;

[0052] Step 32: Average the edge weights between nodes in each subgraph Expands to a scalar repeating vector ( is the subgraph index), as an independent "channel".

[0053] Step 33: Each channel passes through the learnable parameter matrix and Generate query vectors separately and key vector , and then calculate the attention weight by scaling the dot product formula :

[0054]

[0055]

[0056]

[0057] in, and is the parameter matrix to be learned, is the dimension of the mean value matrix;

[0058] Step 34: Obtain the final weight value between fusion nodes through weighted summation:

[0059]

[0060] in, are learnable parameters, is the original edge weight between two nodes in each subgraph;

[0061] Step 35: According to Choose whether to connect the edges. That is, the two nodes are connected, otherwise they are not connected.

[0062] Based on the above discussion, in an optional example of the present application, the dimension of the fused node features is three times the dimension of the subgraph node features after aggregation, which is 33,600 dimensions, and a three-channel channel attention mechanism is required when fusing edges.

[0063] S14: The attention mechanism aggregated features of the fusion graph are input into the MAML model, and the meta-task is used to quickly adapt to the fault diagnosis of the electronic load circuit.

[0064] It should be noted that due to the diversity of electronic load circuit fault types and the scarcity of some fault samples, traditional deep learning requires a large amount of labeled data. Meta-learning accumulates cross-task experience in the meta-training stage through the "learning how to learn" paradigm to achieve rapid identification of small samples of new fault modes. This application uses meta-task rapid adaptation to perform fault diagnosis of electronic load circuits.

[0065] Step 41: Training Task Support set To update the two parameters in the classifier , the update formula is:

[0066]

[0067]

[0068] in, and are the two parameters to be updated, is the learning rate, is the cross entropy loss on the support set;

[0069] Step 42: Utilizing QuerySets To evaluate the updated parameters, the optimization goal of meta-learning is:

[0070]

[0071] in, Cross entropy loss on the query set;

[0072] Step 43: For the new fault diagnosis task, based on its support set data Pre-trained meta-parameters and Perform adaptive optimization and obtain task-specific update parameters through rapid adjustment and , realize the transfer of meta-knowledge to new tasks and customized correction;

[0073] Step 44: Pass the total feature matrix of the fusion graph obtained above through a layer of multi-head graph attention network to obtain Expand into column vector Input classifier ,get , and finally use Calculate the failure probability P:

[0074]

[0075] Among them, K is the number of fault types, and the fault type corresponding to the maximum probability is the diagnosis result.

[0076] Based on the above discussion, in an optional embodiment of the present application, The dimension is 1*3500, and the parameters can be learned , The dimension of the matrix is 700*3500. , It is a 700-dimensional column vector. The training task includes the supervised learning process of known fault modes, and the testing task represents the actual verification process of the model's generalization ability.

[0077] Specific examples are used in this application to illustrate the principles and implementation methods of the present invention. The description of the above examples is only intended to help explain the method and core concept of the present invention. It should be noted that for ordinary people in this technical field, various improvements and modifications can be made to the present invention without departing from the principles of the present invention, and such improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. An electronic load circuit fault diagnosis method based on a multi-channel graph attention network, characterized by: Each vibration sensor signal is evenly divided into several time slices. The frequency domain information of each slice signal is used as a graph node. Edges are connected between graph nodes based on the feature similarity weight coefficients. Each vibration sensor signal constitutes a subgraph, thus establishing a multi-subgraph structure. A multi-head graph attention mechanism is used to aggregate and update the node feature vectors in each subgraph; The updated nodes from the same time slice of different subgraphs are spliced and fused in a head-to-tail manner to form a fusion graph. The edge weight learning mechanism and multi-channel attention mechanism are used to train the edge connection weights between the nodes of the fusion graph. The nodes with the same number are spliced and fused head-to-tail to form a fusion node h. mix ,h mix =||h' ic , (c=1,2,3,…,k), then the edge connection between the fusion nodes is completed through the multi-channel attention mechanism, and the average edge weight m between the nodes of each subgraph is formed into a vector by repeating the scalar value i is the sub-image number, which is treated as an independent "channel", and then the attention weight a of each channel is calculated i , in W1 i ,W2 i is the transformation matrix, d i is the length of the average matrix, and finally the weight value between the fusion nodes is obtained by weighted summation e'=σ(Σa i *W e *e),W e is a learnable parameter, e is the original edge weight between two nodes in each subgraph, and the edge connection is determined based on e'. If e'>0.8, the two nodes are connected, otherwise they are not connected. The attention mechanism aggregated features of the fusion graph are input into the MAML model, and the fault diagnosis of the electronic load circuit is realized based on the meta-learning model.

2. The electronic load circuit fault diagnosis method based on a multi-channel graph attention network according to claim 1, characterized in that: Each vibration sensor signal is evenly divided into several time slices. The frequency domain information of each slice signal is used as a graph node. Edges are connected based on the feature similarity weight coefficients between the graph nodes. Each vibration sensor signal constitutes a subgraph, thus establishing a multi-subgraph structure, including: The sensor signal is cut into several time slices. Take the sample slice X(t0) at a certain time t0 as an example. m is the sample length, k is the number of sensors, and the original sample signals from the same sensor are preprocessed with MAX-MIN normalization and FFT transformation before being used as feature representations for node embedding. In addition, to facilitate the subsequent fusion process, the signals from different sensors at the same time are uniformly labeled. For example, subgraph1.1, subgraph2.1, and subgraph3.1 represent the first time slice nodes in subgraphs 1, 2, and 3, respectively. Then, a learnable parameter W is assigned to each node in the same subgraph. The W between each two nodes is different, and the edge connection is determined based on the learned parameter W'. If W'>0.8, the two nodes are connected, otherwise they are not connected.

3. The electronic load circuit fault diagnosis method based on multi-channel graph attention network according to claim 1, characterized in that: A multi-head graph attention mechanism is used to aggregate and update the node feature vectors in each subgraph, including: For each node x in the subgraph i , we need to first calculate the central node x i Similarity coefficient with its neighbor nodes, x i with x j The similarity coefficient e between ij Defined as, e ij =a(W r h i ||W r h i ),h i and h j Represents node x respectively i and node x j The node features, "||" represents the feature concatenation operation, the matrix W r And mapping a is the parameter to be learned, and then use the obtained e ij As well as the activation function LeakyReLU, the function SOFTMAX calculates the attention coefficient A ij , N i is x i The collection of neighbor nodes, and finally use the attention coefficient to the central node x i Aggregate the node features h' i =σ(∑A ij *W*h j ),σ is the activation function Sigmoid. In order to consider the influence of adjacent nodes on the central node in different aspects, a multi-head graph attention mechanism is used to independently calculate multiple attention heads. The node features calculated by all attention heads are connected to obtain the final aggregated node features. The superscript y represents the parameters of the y-th attention head.

4. The electronic load circuit fault diagnosis method based on a multi-channel graph attention network according to claim 1, characterized in that: The attention mechanism aggregated features of the fusion graph are fed into the MAML model, and meta-tasks are used to quickly adapt to the fault diagnosis of electronic load circuits, including: Using classifier o=W m X all +b m To perform fault diagnosis operations, where W m ,b m is a meta-parameter, which is trained by the inner loop and the outer loop: First, the training task T i The support set S i To update W m and b m , This is the inner loop, and then use the query set Q i To evaluate the updated parameters, the optimization goal of meta-learning is This is the outer loop. Then, for the new fault diagnosis task, use the support set S new To fine-tune the meta-parameter W m , b m , get the updated meta-parameter W new ,b new Then the total feature matrix of the fusion graph is passed through a layer of multi-head graph attention network to obtain X all Input the fully connected layer of the classifier and get o i ,Finally, SOFTMAX is used to calculate the failure probability, The fault type corresponding to the maximum probability is the diagnosis result.

Images