Electronic load circuit abnormity diagnosis method based on two-channel graph convolutional network

Abstract

The invention provides an electronic load circuit abnormity diagnosis method based on a two-channel graph convolutional network. Each component in the electronic load circuit is used as a graph node, the actual connection relation of the components is used as an edge, the electronic load circuit is converted into a graph structure, and feature vectors of the graph nodes are frequency domain information of current and voltage of the components; creating a degree matrix according to the graph structure, determining diagonal elements of the degree matrix by the number of connection edges of each node, creating an adjacent matrix according to the graph structure, and determining values of the adjacent matrix by the connection relationship between the nodes; respectively processing the current and voltage data in two channels by adopting a GCN (Graph Convolutional Neural Network); fusing the feature data obtained by the two channels by adopting a channel attention mechanism; and carrying out anomaly diagnosis by adopting an SVM large-spacing classification model. According to the electronic load circuit abnormity diagnosis method based on the dual-channel graph convolutional network, abnormity detection of each component of the electronic load circuit is realized.

Description

Technical Field

[0001] This application belongs to the fields of machine learning and electronic load, and specifically designs an abnormal diagnosis method for electronic load circuits based on a dual-channel graph convolutional network. Background Art

[0002] Electronic load circuits are widely used in fields such as power electronic equipment and battery testing. However, long-term operation may lead to abnormalities, such as current and voltage fluctuations or component damage. In order to improve stability and reliability, it is crucial to diagnose abnormalities in a timely manner.

[0003] Traditional diagnostic methods rely on expert experience, threshold setting, or signal processing techniques, but their effects are limited in complex environments. Machine learning methods such as SVM and CNN have certain advantages, but it is difficult to capture the topological relationships between components. Graph neural networks (GCNs) are good at processing topological structure data and can improve the accuracy of anomaly detection. This paper proposes a diagnostic method based on a dual-channel graph convolutional network (GCN), which represents the circuit state through a graph structure and uses a dual-channel mechanism to extract the information of current and voltage respectively to enhance the ability to identify abnormalities. Summary of the Invention

[0004] The purpose of this application is to provide an intelligent method to detect abnormalities in each component of an electronic DC load circuit and ensure the timely discovery of faults in the electronic load.

[0005] To solve the above technical problems, this application proposes an abnormal diagnosis method for electronic load circuits based on a dual-channel graph convolutional network, including: Establish a circuit connection information graph according to the connection relationship of circuit components, and each node carries a feature vector; Obtain the degree matrix and adjacency matrix according to the circuit connection information graph, and establish current and voltage feature matrices based on the current and voltage feature vectors corresponding to each node in the circuit connection information graph; Use a graph convolutional neural network to perform convolution on the degree matrix, adjacency matrix, and feature matrix in two channels of voltage and current to obtain two deep feature matrices; Use a channel attention mechanism to fuse the current and voltage deep feature matrices; Use an SVM large margin classification model to perform abnormal diagnosis on the electronic load circuit.

[0006] Optionally, the establishment of the circuit connection information graph according to the connection relationship of circuit components, where each node carries two feature vectors, includes: Sample the current and voltage information of each component, perform spectral transformation on the current and voltage information to obtain frequency-domain information, define each device in the electronic DC load as a node carrying current and voltage frequency-domain information, define the connection relationship between devices as an edge, use the current and voltage frequency-domain information of the device as two feature vectors of the node, and form an electronic load circuit connection information graph with each node carrying the feature vectors based on the edge connection relationship.

[0007] Optionally, obtaining the degree matrix and the adjacency matrix according to the circuit connection information graph, and establishing the current and voltage feature matrix based on the current and voltage feature vectors corresponding to each node in the circuit connection information graph, includes: The degree matrix is formed by converting the sequence of the number of edges connected by each node plus 1 into a diagonal matrix, the adjacency matrix is formed by adding the "0 / 1" matrix representing the connection relationship and the identity matrix of the same dimension, take a number of frequency values at equal step lengths in the frequency domain, and use the current and voltage data of each component corresponding to these frequency values to form the current and voltage feature matrix corresponding to the node.

[0008] Optionally, using the graph convolutional neural network to perform convolution on the degree matrix, the adjacency matrix, and the feature matrix in two channels of voltage and current to obtain two deep feature matrices, includes: Build two parallel graph convolutional networks GCN with two layers, input the obtained degree matrix, adjacency matrix, and current feature matrix into the current channel for feature aggregation calculation to obtain the current deep feature matrix, and input the degree matrix, adjacency matrix, and voltage feature matrix into the voltage channel for feature aggregation calculation to obtain the voltage deep feature matrix.

[0009] Optionally, using the channel attention mechanism to fuse the two deep feature matrices of current and voltage, includes: Perform global average pooling GAP on the current and voltage deep feature matrices to obtain two average feature values, these two values form a column vector, this column vector passes through a two-layer fully connected network to obtain the attention weights corresponding to the current and voltage channels, and finally perform weighted fusion on the two deep feature matrices to obtain the total feature matrix.

[0010] Optionally, using the SVM large margin classification model to perform anomaly diagnosis on the electronic load circuit, includes: According to the training data, train to obtain ten decision functions corresponding to the hyperplanes that can separate normal and abnormal and have the maximum distance from the normal and abnormal data points, input each row of the total feature matrix into the corresponding decision function to obtain ten decision values, and then use the Platt scaling method to predict the anomaly probability.

[0011] This application proposes an abnormal diagnosis method for an electronic load circuit based on a dual-channel graph convolutional network. Each component in the electronic load circuit is used as a graph node, and the actual connection relationship of the components is used as an edge, converting the electronic load circuit into a graph structure. The feature vector of the graph node is the frequency-domain information of the current and voltage of the device; a degree matrix is created according to the graph structure, and the diagonal elements of the degree matrix are determined by the number of connected edges of each node. An adjacency matrix is created according to the graph structure, and its value is determined by the connection relationship between nodes; the graph convolutional neural network GCN is used to process the current and voltage data separately in two channels; a channel attention mechanism is used to fuse the feature data obtained from the two channels; an SVM large margin classification model is used for abnormal diagnosis. This application proposes an abnormal diagnosis method for an electronic load circuit based on a dual-channel graph convolutional network, realizing the abnormal detection of each component in the electronic load circuit. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0013] Figure 1 It is a schematic flow chart of the abnormal diagnosis method for an electronic load circuit based on a dual-channel graph convolutional network provided by an embodiment of this application, realizing the abnormal detection of each component in the electronic load circuit; Figure 2 It is the electronic load circuit provided by an embodiment of the present invention; Figure 3 It is the circuit connection information graph transformed from the circuit diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] In order to enable those skilled in the art to better understand the solution of the present invention, the following will further elaborate on the present invention in conjunction with the drawings and specific embodiments. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0015] Figure 1 It is a schematic flow chart of the abnormal diagnosis method for an electronic load circuit based on a dual-channel graph convolutional network provided by an embodiment of this application, realizing the abnormal detection of each component in the electronic load circuit, specifically including five contents.

[0016] S11: Establish a circuit connection information graph according to the connection relationship of circuit components, and each node carries two feature vectors.

[0017] It should be noted that there are actual physical connection relationships among the components in the electronic load circuit. Based on this connection relationship, the working states of the components affect each other. If a certain part fails, its connected part will surely be interfered. In this application, it is selected to transform the actual circuit topology into a circuit connection information graph for processing.

[0018] Step 11: For Figure 2 the current and voltage information of the ten components, namely R1, R2, R3, R4, programmable DA, MCU, BG1, BG2, BG3, and sampling resistor in the electronic load circuit, in the same time segment, use the fast Fourier transform shown in formula (1) to convert the time-domain data into frequency-domain information;

[0019] Step 12: Define the ten components, namely R1, R2, R3, R4, programmable DA, MCU, BG1, BG2, BG3, and sampling resistor in the Figure 2 electronic load circuit as ten graph nodes;

[0020] Step 13: Average and take n points in the effective frequency range of the current and voltage frequency-domain information of each component, and use the frequency-domain data values of the current and voltage corresponding to each point as the current and voltage feature vectors of the corresponding node of the component;

[0021] Step 14: Define the connection relationships of the components in the Figure 2 electronic load circuit diagram as edges, and use these edges to connect the ten nodes carrying feature vectors to obtain the Figure 3 circuit connection information graph as shown.

[0022] Based on the above discussion, in an optional embodiment of this application, the number of graph nodes is 10, and the dimension n of the current and voltage feature vectors corresponding to each node is 1000 dimensions.

[0023] S12: Obtain the degree matrix and adjacency matrix according to the circuit connection information graph, and establish the current and voltage feature matrices based on the current and voltage feature vectors corresponding to each node in the circuit connection information graph.

[0024] It should be noted that the data in the circuit connection information graph needs to be represented in a digital manner before it can be operated. In this application, it is selected to obtain the degree matrix and adjacency matrix representing its features according to the circuit connection information graph, and establish the current and voltage feature matrices from the feature vectors obtained by collecting data.

[0025] Step 21: For Figure 3The ten nodes in the circuit connection information diagram are numbered from 0 to 9. From the diagram, the number of edges connected to each of the ten nodes can be obtained. Arrange them in the order from 0 to 9 as a sequence, and add 1 to each item respectively.

[0026] Step 22: Convert this sequence into a diagonal matrix, and this diagonal matrix is the degree matrix of the circuit connection information diagram. ;

[0027] Step 23: According to Figure 3 the circuit connection information diagram and the numbers of each node, use number pairs to represent the connection relationships of each node, and obtain the same number of number pairs as the number of edges.

[0028] Step 24: Construct a 10*10 zero matrix, whose rows and columns respectively represent the ten nodes from 0 to 9. Assign the value 1 to the positions corresponding to the constructed number pairs in this matrix. Swap the two numbers of all number pairs to get a new set of number pairs, and perform the same assignment operation as above. Add the matrices obtained from the two assignments to an identity matrix of the same dimension as it to obtain the final adjacency matrix. ;

[0029] Step 25: Concatenate the current feature vectors of the ten nodes from 0 to 9 from top to bottom into a 10*n matrix to form the current feature matrix of the circuit connection information diagram. , and similarly obtain the voltage feature matrix. ;

[0030] Based on the above discussion, in an optional embodiment of the present application, a total of four matrices are obtained. Matrix and are both 10*10 square matrices, and are 10*1000 matrices.

[0031] S13: Use a graph convolutional neural network to perform convolution on the degree matrix, adjacency matrix, and feature matrix in two channels of voltage and current to obtain two deep feature matrices.

[0032] It should be noted that the working state of the components has different effects on their current and voltage signals. The present application chooses to process the current and voltage information in two channels.

[0033] Step 31: According to the degree matrix and the adjacency matrix , calculate and obtain ;

[0034] Step 32: The current feature matrix and the voltage feature matrix Two GCN channels are respectively input for calculation, as shown in formulas (3)(4) and (5)(6), to obtain two depth feature matrices of current and voltage Denote it as ;

[0035] Based on the above discussion, in an optional embodiment of the present application, the parameter matrix has a dimension of 1000*700, and the parameter matrix has a dimension of 700*300. The finally obtained depth feature matrix is a matrix of 10*300, is the activation function.

[0036] S14: Use the channel attention mechanism to fuse the two depth feature matrices of current and voltage.

[0037] It should be noted that the depth feature matrices obtained after processing the current and voltage information contribute differently to the final anomaly diagnosis. Therefore, a set of weights is required to process and fuse the two. The present application selects the channel attention mechanism to fuse the depth feature matrices of current and voltage.

[0038] Step 41: As shown in formulas (7)(8), perform global average pooling (GAP) on the depth feature matrices of current and voltage, are the two dimensions of the depth feature matrix respectively, to obtain two values and form them into a two-dimensional column vector , as shown in formula (9);

[0039] Step 42: As shown in formulas (10)(11), input the vector into a two-layer fully connected network for calculation to obtain a two-dimensional column vector , which represents the weight vector of the channel attention weights of current and voltage, as shown in formula (12);

[0040] Step 43: As shown in formula (13), according to the channel attention weight vector Among the two weight values, the current and voltage depth feature matrices are fused to obtain the total feature matrix;

[0041] Based on the above discussion, in an optional embodiment of the present application, are 10 and 300 respectively, and the learnable weights of the two-layer fully connected network have a dimension of 1*10, and the weights have a dimension of 10*1, is the activation function.

[0042] S15: An SVM large margin classification model is used to perform anomaly diagnosis on the electronic load circuit.

[0043] It should be noted that the anomaly detection of each component in the electronic load circuit is essentially close to a binary classification problem. Based on the total feature matrix for classification, the present application selects an SVM large margin classification model to calculate the classification probability.

[0044] Step 51: For this non-linear binary classification problem of anomaly detection, an RBF kernel function is introduced as shown in formula (14);

[0045] Step 52: The ten rows of data in the total feature matrix are split and used to train ten sets of decision function parameters for classifying ten components respectively. The parameters include the Lagrange multipliers of the support vectors , the labels of the support vectors , the support vectors , the bias term , and ten different parameters are obtained, in the form of the decision function in formula (15);

[0046] Step 53: Based on the value obtained from the decision function , the Platt Scaling method shown in formula (16) is used for anomaly probability estimation, are parameters obtained by fitting from the training set through the maximum likelihood method. This method outputs a probability between 0 and 1 for each decision function value, representing the anomaly probability of the corresponding component;

[0047] Based on the above discussion, in an optional example of the present application, the dimension of the support vector is the same as the dimension of the total feature matrix, which is 300 dimensions, and the other parameters in each group are also 300;

[0048] In this application, specific examples are used to illustrate the principles and implementation manners of the present invention. The description of the above embodiments is only for helping to understand the method and its core idea of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. An electronic load circuit abnormality diagnosis method based on a dual-channel graph convolutional network is characterized by: A circuit connection information graph is established based on the connection relationship of circuit components, and each node carries two feature vectors; Obtain a degree matrix and an adjacency matrix according to the circuit connection information graph, and establish a current and voltage feature matrix based on the current and voltage feature vectors corresponding to each node in the circuit connection information graph; The graph convolutional neural network is used to convolve the degree matrix, adjacency matrix and feature matrix in two channels: voltage and current, to obtain two deep feature matrices. The channel attention mechanism is used to fuse the two deep feature matrices of current and voltage; The SVM large-margin classification model is used to perform abnormal diagnosis on electronic load circuits.

2. The electronic load circuit abnormality diagnosis method based on a dual-channel graph convolutional network according to claim 1, characterized in that: A circuit connection information graph is established based on the connection relationship of circuit components. Each node carries two feature vectors, including: The current and voltage information of each component is sampled, and the current and voltage information are transformed into frequency domain information by spectrum transformation. Each component in the electronic DC load is defined as a node carrying current and voltage frequency domain information, and the connection relationship between components is defined as an edge. The current and voltage frequency domain information of the device is used as the two eigenvectors of the node, and the nodes carrying the eigenvectors are combined into an electronic load circuit connection information graph based on the connection relationship of the edges.

3. The electronic load circuit abnormality diagnosis method based on dual-channel graph convolutional network according to claim 1, characterized in that: The degree matrix and adjacency matrix are obtained according to the circuit connection information graph, and the current and voltage feature matrices are established based on the current and voltage feature vectors corresponding to each node in the circuit connection information graph, including: The degree matrix is ​​composed of a sequence consisting of the number of edges connected to each node plus 1 converted into a diagonal matrix. The adjacency matrix is ​​composed of a "0\1" matrix representing the connection relationship plus a unit matrix of the same dimension. Several frequency values ​​are taken with equal steps in the frequency domain, and the current and voltage data of each component corresponding to these frequency values ​​are used to form the current and voltage characteristic matrix corresponding to the node.

4. The electronic load circuit abnormality diagnosis method based on a dual-channel graph convolutional network according to claim 1, characterized in that: The graph convolutional neural network is used to convolve the degree matrix, adjacency matrix and feature matrix in two channels: voltage and current, to obtain two deep feature matrices, including: Two parallel two-layer graph convolutional networks (GCNs) are built, and the degree matrix, adjacency matrix, and current feature matrix are input into the current channel for feature aggregation calculation to obtain the current deep feature matrix. The degree matrix, adjacency matrix, and voltage feature matrix are input into the voltage channel for feature aggregation calculation to obtain the voltage deep feature matrix.

5. The electronic load circuit abnormality diagnosis method based on dual-channel graph convolutional network according to claim 1, characterized in that: The channel attention mechanism is used to fuse the two deep feature matrices of current and voltage, including: The current and voltage deep feature matrices are subjected to global average pooling GAP to obtain two average eigenvalues. These two values ​​​​compose a column vector. This column vector is passed through a two-layer fully connected network to obtain the attention weights corresponding to the current and voltage channels. Finally, the two deep feature matrices are weightedly fused to obtain the total feature matrix.

6. The electronic load circuit abnormality diagnosis method based on a dual-channel graph convolutional network according to claim 1, characterized in that: The SVM large-interval classification model is used to perform abnormal diagnosis of electronic load circuits, including: According to the training data, ten decision functions corresponding to the hyperplanes that can separate normal and abnormal data points and have the largest distance from normal and abnormal data points are obtained through training. Each row of the total feature matrix is ​​input into the corresponding decision function to obtain ten decision values, and then the Platt scaling method is used to predict the abnormal probability.

Images