A Fault Diagnosis Method for Analog Circuits Based on Graph Neural Network

Abstract

The present invention discloses a method for fault diagnosis of analog circuits based on a graph neural network. When constructing a fault sample graph, the method forms constraints between structures with structural features between samples, that is, edges between sample nodes; when the graph neural network processes the fault sample graph, the structural features and data features of the samples are extracted to classify the fault states of the samples. This structural constraint in the fault sample graph not only helps the model extract more effective features for fault classification, thereby improving the fault diagnosis accuracy of the model; it also plays the role of sample clustering, so that the model also has a good training effect in the few-sample learning with a low proportion of training set samples.

Description

Technical Field

[0001] The present invention relates to the technical field of analog circuit fault diagnosis, and more specifically to an analog circuit fault diagnosis method based on graph neural network. Background Art

[0002] In typical electronic equipment, compared with digital circuits, analog circuit fault diagnosis is more difficult and has a higher fault incidence rate. Analog circuits, which account for only 20% of the scale, may have more than 80% of faults. The fault diagnosis technology of analog circuits restricts the development of the fault diagnosis technology of the entire electronic system. Analog circuit faults can be divided into soft faults and hard faults according to their degree. Soft faults refer to faults caused by the deviation of the parameters of circuit components beyond the predetermined tolerance range, which will cause abnormal equipment performance while the circuit can still work. Hard faults refer to faults caused by a sudden and excessive change in a parameter of a circuit component, which will cause sudden paralysis of electronic equipment and cause significant losses. When a soft fault occurs, the fault characteristics of the circuit are not as obvious as when a hard fault occurs, which makes it more difficult to diagnose a soft fault. However, in order to avoid serious losses, the fault should still be diagnosed in time at the soft fault stage. Improving the accuracy of analog circuit fault diagnosis, especially the accuracy of analog circuit soft fault diagnosis, is crucial to improving the overall reliability of electronic equipment.

[0003] The fault diagnosis method of analog circuits has gone through three development stages: manual observation method, traditional diagnosis method based on complex models, and intelligent diagnosis method based on data-driven. As the scale of analog circuits in the circuit system of equipment becomes larger and larger and the structure becomes more and more complex, the difficulty of diagnosing them is also increasing. The fault diagnosis methods of the first two stages are no longer applicable to most circuits. The intelligent diagnosis method based on data-driven is combined with machine learning, neural networks, and other artificial intelligence technologies. It can distinguish different types of fault states without establishing complex diagnostic models through circuit topology and known parameters, which has great advantages in the field of analog circuit fault diagnosis. In recent years, the continuous development of data information processing technologies such as wavelet transform preprocessing, one-dimensional convolutional neural networks, deep belief networks, and adaptive neural fuzzy systems has also provided continuous technical support for data-driven intelligent diagnosis methods. Although the data-driven intelligent diagnosis method combined with artificial intelligence technology has been the mainstream research in the field of fault diagnosis since its development, it is often limited by the principles of general artificial intelligence technology itself, and can only process fault sample data with regular data structure, thereby ignoring the structural characteristics between samples; in order to retain the structural characteristics between samples, the fault sample set can be constructed as a sample arrangement structure and irregular graph structure data. Graph neural network is a kind of neural network that is directly defined to perform calculations on graph structured data. Graph convolutional neural network is one of the variant models of graph neural network. It completes a graph convolution calculation by aggregating the node features of neighboring nodes and the central node itself, updating the features of the central node, and stacking multiple layers of graph convolutional networks to extract the data features of samples in the fault sample graph and the structural features between samples, which can effectively improve the fault diagnosis accuracy of the analog circuit fault diagnosis model.

[0004] Therefore, how to provide an analog circuit fault diagnosis method based on graph neural network with high accuracy is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] In view of this, the present invention provides an analog circuit fault diagnosis method based on graph neural network.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A method for diagnosing analog circuit faults based on graph neural network, comprising the following steps:

[0008] S1. The pulse response signal at the output of the sampling circuit is used to construct a fault sample data set, where each type of fault state includes at least one set of sample data;

[0009] S2. Dividing the fault sample data set into a training set and a test set;

[0010] S3. Take the sampled samples as nodes V, and the multidimensional data features of the samples as node features to form a node feature matrix X, and calculate the cosine similarity between the nodes; the connection relationship E between each node and a fixed number of most similar neighboring nodes is used as an edge to construct a fault sample graph, and the fixed number is determined by the connection accuracy between the training set sample nodes of the same category in the fault sample graph;

[0011] S4. Initialize the parameters of the graph neural network. In the training phase, all nodes participate in the forward transmission of node features, and only calculate the loss function and optimize and update the corresponding parameters on the nodes of the training set. The above process is repeated until the model converges, and a fault diagnosis classification model based on the graph neural network is established.

[0012] S5. Use the fault diagnosis classification model based on graph neural network to predict the category labels of the test set nodes and output the diagnosis results.

[0013] Preferably, the specific content of S3 includes:

[0014] S31. Select the fault sample as a node, the sample data as the node feature, and the feature dimension as N; build the connection relationship between the nodes according to the node features:

[0015] For any two samples x in the fault sample set i and i , where i = 1, 2, ..., N, calculate two samples x i and i The method of cosine similarity r is:

[0016]

[0017] Among them, cosine similarity is defined as the cosine value of the angle between sample vectors; the closer the absolute value is to 1, the closer the angle between sample vectors is to 0, indicating that the two samples are more similar;

[0018] S32. Circularly calculate the cosine similarity of each node with the remaining nodes, and consider that each node and a fixed number of nodes that are most similar belong to the same fault state, and the connection relationship between the nodes in the same fault state is regarded as an edge;

[0019] S33. After determining the connection relationship between nodes, a fault sample graph is constructed.

[0020] Preferably, the specific content of the forward transmission of the node characteristics in S4 includes:

[0021] S411. Input the fault sample graph into the graph neural network, and the connection relationship E of the nodes of the fault sample graph is represented by an adjacency matrix A:

[0022]

[0023] Where N represents the number of samples;

[0024] S412. Update node features based on the fault diagnosis classification model of the graph neural network, including:

[0025] (1) Complete the node feature update, where the node feature matrix of the l+1th layer is:

[0026] H (l+1) =σ(AH (l) W (l) )

[0027] Among them, H (l) is the node feature matrix of the lth layer, W (l) is the weight matrix of the lth layer, the final output matrix is ​​Z, and σ(·) represents the nonlinear activation function;

[0028] (2) Obtain the matrix of the aggregate node’s own features based on the adjacency matrix A

[0029]

[0030] (3) By aggregating the matrix of node features Update node features, including:

[0031] Obtain the degree matrix D, which is a diagonal matrix. The elements on the main diagonal represent the degrees of the corresponding nodes. Based on the degree matrix D, add the main diagonal elements to the degrees of the corresponding nodes to obtain use right Normalize to get

[0032]

[0033]

[0034] according to The graph convolutional network layer is improved. The improved graph convolutional network layer is:

[0035]

[0036] Stack the graph convolutional network layers to obtain the implicit expression of node features and send them to the downstream classification layer.

[0037] Preferably, in S411, the hyperparameters of the graph neural network are initialized before the fault sample graph is input into the graph neural network, and the hyperparameters include the number of graph convolutional neural network layers, the node feature dimension of the hidden layer, the learning rate and the number of iterations.

[0038] Preferably, the specific contents of calculating the loss function and optimizing and updating the corresponding parameters on the training set nodes in S4 include:

[0039] S421. Select an activation function; the activation function between graph convolutional network layers is the relu function, and the activation function between the graph neural network layer and the classification layer is the log (softmax) function, where:

[0040] relu(x)=max(0,x)

[0041]

[0042] The log(soffmax) function maps the final implicit expression of the node feature to the probability interval of (-∞, 0), and selects the category with the highest probability as the classification result of the corresponding sample node;

[0043] S422. Select the cross entropy function as the loss function. If the true category of sample i is c, then y ic Take 1, otherwise y ic Take 0; p ic is the predicted probability that sample i belongs to category c;

[0044]

[0045] Where N is the number of categories, y ic is a symbolic function;

[0046] S423. Select Adam algorithm as the optimization function, m i and v i The update is:

[0047] m i =β1m i-1 +(1-β1)g i

[0048]

[0049] Among them, m i is the exponential moving mean of the gradient, obtained by the first-order moment of the gradient; v i is the square gradient, obtained by the second-order moment of the gradient; i represents the number of iterations; β1 and β2 are constants that control exponential decay; g i is the first-order derivative;

[0050] S424. Acquisition and in For mi's correction, v i Correction:

[0051]

[0052]

[0053] S425. Obtain weight matrix parameter w i Update, where α is the learning rate; ε is a constant to maintain numerical stability:

[0054]

[0055] Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses an analog circuit fault diagnosis method based on a graph neural network. When constructing a fault sample graph, the method forms structural features between samples into constraints between structures, that is, edges between sample nodes; when the graph neural network processes the fault sample graph, the structural features and data features of the samples are extracted to classify the fault states of the samples. This structural constraint in the fault sample graph not only helps the model extract more effective features for fault classification, thereby improving the fault diagnosis accuracy of the model; it also plays the role of sample clustering, so that the model also has a good training effect in the few-sample learning with a low proportion of training set samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0057] Figure 1 A schematic diagram of a flow chart provided in a method for diagnosing faults in analog circuits based on a graph neural network according to the present invention;

[0058] Figure 2 This is a schematic diagram of a four-op-amp dual-second-order filter circuit provided in an embodiment of an analog circuit fault diagnosis method based on a graph neural network of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The embodiment of the present invention discloses a method for diagnosing analog circuit faults based on a graph neural network, such as Figure 1 As shown, the following steps are included:

[0061] S1. The pulse response signal at the output of the sampling circuit is used to construct a fault sample data set, where each type of fault state includes at least one set of sample data;

[0062] S2. Divide the fault sample data set into a training set and a test set;

[0063] S3. Take the sampled samples as nodes V and the multidimensional data features of the samples as node features to form a node feature matrix X, and calculate the cosine similarity between the nodes. The connection relationship E between each node and a fixed number of most similar neighboring nodes is used as an edge to construct a fault sample graph. The fixed number is determined by the connection accuracy between the training set sample nodes of the same category in the fault sample graph.

[0064] S4. Initialize the parameters of the graph neural network. In the training phase, all nodes participate in the forward transmission of node features, and only calculate the loss function and optimize and update the corresponding parameters on the nodes of the training set. Train until the model converges, and establish a fault diagnosis classification model based on the graph neural network.

[0065] S5. Use the fault diagnosis classification model based on graph neural network to predict the category labels of the test set nodes and output the diagnosis results.

[0066] It should be noted that:

[0067] In S3, the closer the cosine similarity between nodes is to 1, the more similar the features of the two nodes are, and the more likely the two nodes belong to the same category; a fixed number of the most similar neighbor nodes of each node are screened out, and a connection relationship is established as the edge E, and the fault sample graph is composed of the mutual connections between nodes of the same category identified by the cosine similarity formula; wherein, if the sample nodes under each fault state have a connection relationship with the sample nodes under the same state, the connection is accurate; the fixed number is determined by the average connection accuracy between the training set sample nodes of all categories in the fault sample graph.

[0068] Calculate the percentage of correct connection relationships under each fault state, which is the connection accuracy under this state; take the average value of the clustering accuracy of all states, which is the clustering accuracy of the constructed fault sample graph.

[0069] In order to further implement the above technical solutions, the specific contents of S3 include:

[0070] S31. Select the fault sample as a node, the sample data as the node feature, and the feature dimension as N; build the connection relationship between the nodes according to the node features:

[0071] For any two samples x in the fault sample set i and i , where i = 1, 2, ..., N, calculate two samples x i and i The method of cosine similarity r is:

[0072]

[0073] Among them, cosine similarity is defined as the cosine value of the angle between sample vectors; the closer the absolute value is to 1, the closer the angle between sample vectors is to 0, indicating that the two samples are more similar;

[0074] S32. Circularly calculate the cosine similarity of each node with the remaining nodes, and consider that each node and a fixed number of nodes that are most similar belong to the same fault state, and the connection relationship between the nodes in the same fault state is regarded as an edge;

[0075] S33. After determining the connection relationship between nodes, a fault sample graph is constructed.

[0076] In order to further implement the above technical solution, the specific content of the forward transmission of node features in S4 includes:

[0077] S411. Input the fault sample graph into the graph neural network, and the connection relationship E of the nodes in the fault sample graph is represented by the adjacency matrix A:

[0078]

[0079] Where N represents the number of samples;

[0080] S412. In the graph convolutional neural network, the aggregation and update message transmission mechanism is used to complete the update of node features. The central node updates the central node features by aggregating the node features of neighboring nodes and its own node features. The fault diagnosis classification model based on the graph neural network updates the node features, and the specific contents include:

[0081] (1) Complete the node feature update, where the node feature matrix of the l+1th layer is:

[0082] H (l+1) =σ(AH (l) W (l) )

[0083] Among them, H (l) is the node feature matrix of the lth layer, W (l) is the weight matrix of the lth layer, the final output matrix is ​​Z, and σ(·) represents the nonlinear activation function;

[0084] (2) Since the diagonal elements of the adjacency matrix are 0, the above formula does not have the characteristics of the aggregation center node itself when performing graph convolution. Therefore, the unit matrix I is added to the adjacency matrix A. N ,get

[0085]

[0086] (3) By aggregating the matrix of node features When the node features are updated, the specific contents include:

[0087] Get the degree matrix D. The degree matrix D is a diagonal matrix. The elements on the main diagonal represent the degrees of the corresponding nodes. Based on the degree matrix D, add the main diagonal elements to the degrees of the corresponding nodes and get use right Normalize to get

[0088]

[0089]

[0090] according to The graph convolutional network layer is improved. The improved graph convolutional network layer is:

[0091]

[0092] Stack the graph convolutional network layers to obtain the implicit expression of node features and send them to the downstream classification layer.

[0093] Stacking multiple layers of graph convolution allows the central node to learn deeper node features.

[0094] In order to further implement the above technical solution, in S411, the hyperparameters of the graph neural network are initialized before the fault sample graph is input into the graph neural network. The hyperparameters include the number of graph convolutional neural network layers, the node feature dimension of the hidden layer, the learning rate and the number of iterations.

[0095] In order to further implement the above technical solution, the specific contents of calculating the loss function and optimizing and updating the corresponding parameters on the training set node in S4 include:

[0096] S421. Select a suitable activation function to enable the model to learn more complex functions while improving the computational efficiency of the network. The activation function between graph convolutional network layers is the relu function, and the activation function between the graph neural network layer and the classification layer is the log (softmax) function, which can effectively solve the upper and lower overflow problems of the softmax function, where:

[0097] relu(x)=max(0,x)

[0098]

[0099] The log(softmax) function maps the final implicit expression of the node feature to the probability interval of (-∞, 0), and selects the category with the highest probability as the classification result of the corresponding sample node;

[0100] S422. Select the cross entropy function as the loss function. If the true category of sample i is c, then y ic Take 1, otherwise y ic Take 0; p ic is the predicted probability that sample i belongs to category c;

[0101]

[0102] Where N is the number of categories, y ic is a symbolic function;

[0103] S423. Select Adam algorithm as the optimization function, m i and v i The update is:

[0104] m i =β1m i-1 +(1-β1)g i

[0105]

[0106] Among them, m i is the exponential moving mean of the gradient, obtained by the first-order moment of the gradient; v i is the square gradient, obtained by the second-order moment of the gradient; i represents the number of iterations; β1 and β2 are constants that control exponential decay; g i is the first-order derivative;

[0107] S424. Acquisition and in For mi's correction, v i Correction:

[0108]

[0109]

[0110] S425. Obtain weight matrix parameter w i Update, where α is the learning rate; ε is a constant to maintain numerical stability:

[0111]

[0112] The default values ​​of all the above parameters are set to: α = 0.001, β1 = 0.9, β2 = 0.999, ε = 10 -8 . The smaller the value, the less certain the current direction is, so the smaller the step size is.

[0113] The above contents will be described below through specific embodiments:

[0114] In this embodiment, a four-op-amp biquad filter circuit in the international standard circuit is selected as the experimental circuit.

[0115] (1-1) Use PSPICE software to simulate the four-op-amp biquad filter circuit, use a 10μs pulse signal with an amplitude of 5V and a frequency of 2kHz as the excitation source, and select the circuit output as the test point;

[0116] (1-2) The experimental circuit schematic diagram is as follows Figure 2 As shown, the tolerance ranges of the resistor and capacitor components in the circuit are set to 5% and 10% respectively. Through sensitivity analysis of circuit components, C1, C2, R1, R2, and R3 are selected as the components to be tested. When the component parameter value deviates from the nominal value by about 30%, the circuit is considered to have a soft fault. The fault category, nominal value, and fault value of each component in the experimental circuit are shown in the following table:

[0117]

[0118] (1-3) Through Monte-Carlo analysis, 300 groups of impulse response sample signals are generated for each fault category; each group of signals is sampled with 1000 points, that is, each sample has 1000-dimensional feature data;

[0119] (2) Randomly add training set or test set Mask to samples under each fault state in the ratio of 7:3, 5:5, 3:7, and 1:9;

[0120] (3-1) Select the fault sample as the node and the sample data as the node feature. For any two samples x in the fault sample set, i and i (i=1,2,…,N), N=1000, calculate the cosine similarity;

[0121] (3-2) Calculate the cosine similarity between each node and the remaining nodes in a loop, and select the five most similar nodes for each node to build edge connections. Too many neighbor nodes will introduce incorrect connection relationships, and too few neighbor nodes will lead to insufficient structural constraints, resulting in low generalization ability and computing speed of the model. After experimental verification, in this circuit, the number of most similar nodes for each central node is set to 5, so each node has at least 5 neighbor nodes;

[0122] (3-3) Determine the connection relationship between nodes and complete the construction of the fault sample graph;

[0123] (4-1-1) Initialize the parameters of the graph neural network, set two layers of graph convolution; the hidden layer node feature is 100, the learning rate lr is 0.001, and the number of iterations is 1000;

[0124] (4-1-2) Input the fault sample graph (including sample nodes V, connection relationships E between nodes, and node feature matrix X) into the graph convolutional neural network, where the connection relationship E of the nodes in the graph is usually represented by the adjacency matrix A;

[0125] (4-2-1) In the graph convolutional neural network, the aggregation and update message passing mechanism is used to complete the update of node features. The central node updates the central node features by aggregating the node features of neighboring nodes and its own node features.

[0126] (4-2-2) Stack two layers of graph convolutional networks so that nodes can learn deeper node features during training;

[0127] (4-3-1) The activation function selects the relu function between the graph convolutional network layers, and the log (softmax) function is selected between the graph neural network layer and the classification layer;

[0128] Let the input node feature be X, the adjacency matrix be A, W (l) is the weight matrix of the lth layer, and the final output matrix is ​​Z. The forward update formula of the node features of the two-layer graph convolutional neural network is as follows;

[0129]

[0130] (4-3-2) The loss function selects the cross entropy function;

[0131] (4-3-4) Select Adam algorithm for optimization function;

[0132] (4-3-5) Circularly train the graph neural network until the model converges;

[0133] (5) Use the trained fault diagnosis classification model to predict the category labels of the sample nodes in the test set, and output the diagnosis results and the fault diagnosis accuracy of the model.

[0134] Table 1 shows the test set diagnosis accuracy of the simulated circuit fault diagnosis model under different ratios of training set and test set.

[0135]

[0136] According to the experimental results of the embodiments, the present invention has the following advantages:

[0137] (1) The fault sample set is used to construct a fault sample graph, which retains the structural constraints between samples and gives it the characteristics and advantages of semi-supervised learning;

[0138] (2) Under four different ratios of training set and test set samples, the graph convolutional neural network is used to extract the data features and structural features of the fault sample graph, which can effectively improve the fault diagnosis accuracy compared with only extracting the data features or structural features of the samples;

[0139] (3) Applying graph neural networks to the field of analog circuit fault diagnosis provides new ideas for solving the problem of analog circuit fault diagnosis as the theoretical system of graph neural networks and their variant technologies become more and more mature.

[0140] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0141] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for fault diagnosis of analog circuits based on graph neural networks, characterized in that: The following steps are involved: S1. The pulse response signal at the output of the sampling circuit is used to construct a fault sample data set, where each type of fault state includes at least one set of sample data; S2. Dividing the fault sample data set into a training set and a test set; S3. Take the sampled samples as nodes V, and the multidimensional data features of the samples as node features to form a node feature matrix X, and calculate the cosine similarity between the nodes; the connection relationship E between each node and a fixed number of most similar neighboring nodes is used as an edge to construct a fault sample graph, and the fixed number is determined by the connection accuracy between the training set sample nodes of the same category in the fault sample graph; calculate the percentage of correct connection relationships under each fault state, which is the connection accuracy under this state; S4. Initialize the parameters of the graph neural network. In the training phase, all nodes participate in the forward transmission of node features, and only calculate the loss function and optimize and update the corresponding parameters on the nodes of the training set. Train until the model converges, and establish a fault diagnosis classification model based on the graph neural network. The fault sample graph is input into the graph neural network, and the connection relationship E of the nodes in the fault sample graph is represented by the adjacency matrix A: Where N represents the number of samples; S5. Use the fault diagnosis classification model based on graph neural network to predict the category labels of the test set nodes and output the diagnosis results.

2. The analog circuit fault diagnosis method based on graph neural network according to claim 1 is characterized in that: The specific contents of S3 include: S31. Select the fault sample as a node, the sample data as the node feature, and the feature dimension as N; build the connection relationship between the nodes according to the node features: For any two samples x in the fault sample set i and i , where i = 1, 2, ..., N, calculate two samples x i and i The method of cosine similarity r is: Among them, cosine similarity is defined as the cosine value of the angle between sample vectors; the closer the absolute value is to 1, the closer the angle between sample vectors is to 0, indicating that the two samples are more similar; S32. Circularly calculate the cosine similarity of each node with the remaining nodes, and consider that each node and a fixed number of nodes that are most similar belong to the same fault state, and the connection relationship between the nodes in the same fault state is regarded as an edge; S33. After determining the connection relationship between nodes, a fault sample graph is constructed.

3. The analog circuit fault diagnosis method based on graph neural network according to claim 1 is characterized in that: The specific contents of the forward transmission of node features in S4 include: The fault diagnosis classification model based on graph neural network updates the node features, including: (1) Complete the node feature update, where the node feature matrix of the l+1th layer is: H (l+1) =σ(AH (l) W (l) ) Among them, H (l) is the node feature matrix of the lth layer, W (l) is the weight matrix of the lth layer, the final output matrix is ​​Z, and σ(·) represents the nonlinear activation function; (2) Obtain the matrix of the aggregate node’s own features based on the adjacency matrix A (3) By aggregating the matrix of node features Update node features, including: Obtain the degree matrix D, which is a diagonal matrix. The elements on the main diagonal represent the degrees of the corresponding nodes. Based on the degree matrix D, add the main diagonal elements to the degrees of the corresponding nodes and obtain use right Normalize to get according to The graph convolutional network layer is improved. The improved graph convolutional network layer is: Stack the graph convolutional network layers to obtain the implicit expression of node features and send them to the downstream classification layer.

4. The analog circuit fault diagnosis method based on graph neural network according to claim 3 is characterized in that: In S411, before the fault sample graph is input into the graph neural network, the hyperparameters of the graph neural network are initialized, and the hyperparameters include the number of graph convolutional neural network layers, the node feature dimension of the hidden layer, the learning rate and the number of iterations.

5. The analog circuit fault diagnosis method based on graph neural network according to claim 1 is characterized in that: The specific contents of calculating the loss function and optimizing and updating the corresponding parameters on the training set nodes in S4 include: S421. Select an activation function; the activation function between graph convolutional network layers is the relu function, and the activation function between the graph neural network layer and the classification layer is the log (softmax) function, where: relu(x)=max(0,x) The log(softmax) function maps the final implicit expression of the node feature to the probability interval of (-∞,0), and selects the category with the highest probability as the classification result of the corresponding sample node; S422. Select the cross entropy function as the loss function. If the true category of sample i is c, then y ic Take 1, otherwise y ic Take 0; p ic is the predicted probability that sample i belongs to category c; Where N is the number of categories, y ic is a symbolic function; S423. Select Adam algorithm as the optimization function, m i and v i The update is: m i =β1m i-1 +(1-β1)g i Among them, m i is the exponential moving mean of the gradient, obtained by the first-order moment of the gradient; v i is the square gradient, obtained by the second-order moment of the gradient; i represents the number of iterations; β1 and β2 are constants that control exponential decay; g i is the first-order derivative; S424. Acquisition and in is m i Correction, v i Correction: S425. Obtain weight matrix parameter w i Update; where α is the learning rate; ε is a constant to maintain numerical stability:

Images