Electronic component fault detection method based on artificial intelligence
By constructing a GCNs-LSTM detection model based on graph theory data and optimizing its parameters, the problems of limitations in the existing technology of fault detection results and low detection accuracy are solved, and faster and more accurate fault detection of electronic components are achieved.
Patent Information
- Application Number
- CN202411647873.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-18
- Publication Date
- 2025-03-11
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the prior art, no correlation fault detection is performed through graph theory, resulting in limited detection results and insufficient parameter optimization of deep learning network models, resulting in slow detection speed and low detection accuracy.
A method for detecting faults of electronic components based on artificial intelligence is proposed. By obtaining simulation data and actual measurement data of preset fault types, preprocessing is carried out to construct single graph theory data and associated graph theory data, building GCNs-LSTM detection model, and optimizing model parameters using improved whale optimization algorithm.
The detection speed and detection accuracy of the deep learning network model are improved, and the model output features with nonlinear characteristics are enhanced, which enhances the accuracy of detection of various fault states.
Smart Images
Figure CN119598255A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of electronic components, and specifically refers to an electronic component fault detection method based on artificial intelligence. Background Art
[0002] With the rapid development of the electronics industry, the complexity and diversity of electronic components are constantly increasing. Traditional manual inspection methods can no longer meet the efficiency and accuracy requirements of modern industrial production. In existing technologies, electronic component fault detection usually relies on manual inspection, but this method is costly, inefficient, and the detection results are greatly affected by subjective factors.
[0003] The Chinese invention patent with patent publication number CN117830309A discloses an artificial intelligence-based electronic component fault diagnosis system, which relates to the field of electronic component technology; the system includes: a database, a feature extraction module and a fault diagnosis module; the feature extraction module automatically extracts the physical characteristics and electrical characteristic parameters of the electronic components and uses them for fault diagnosis and performance analysis, which helps to improve detection efficiency and accuracy, reduce fault risks, and provide reliable data support for the maintenance and management of electronic components; the fault diagnosis module obtains physical evaluation index, electrical evaluation index and operating temperature range through physical performance analysis and electrical performance analysis, and determines the fault type of the electronic component based on this, and finally displays the fault results; it realizes automated feature extraction and intelligent fault diagnosis of electronic components, helps to improve maintenance efficiency and reduce fault risks, and provides intuitive fault information through the fault display function, further improving the effect of fault diagnosis.
[0004] Currently, in the field of electronic component fault detection, physical and electrical performance analysis is obtained through specific signal processing methods. Convolutional networks and their derivatives are then used to extract physical and electrical features and perform classification tasks to complete fault detection. Without the use of graph theory for correlated fault detection, the detection results are limited. However, because the performance of deep learning network models is affected by various parameters, setting these parameters to optimize model performance requires extensive experimentation and adjustment. Therefore, optimizing the parameters of deep learning network models to improve detection speed and accuracy is an urgent issue. Summary of the Invention
[0005] In order to solve the problems in the above-mentioned prior art that no associated fault detection is performed through graph theory, the detection results have certain limitations, and the parameters of the deep learning network model are not optimized, resulting in slow detection speed and low detection accuracy, the present invention proposes an electronic component fault detection method based on artificial intelligence to improve the above-mentioned problems.
[0006] The specific application is as follows:
[0007] An artificial intelligence-based electronic component fault detection method, the method comprising:
[0008] S1: Acquire simulation data of preset fault types, where the preset fault types are divided into single fault and combined fault;
[0009] S2: Obtaining actual measured data of the preset fault type;
[0010] S3: Preprocessing the simulation data and the actual measured data to obtain feature data and correlation relationships between the feature data, constructing single graph data based on the feature data, and constructing correlation graph data based on the correlation relationships between the feature data;
[0011] S4: constructing an electronic component detection model, and training the electronic component detection model based on the single graph theory data and the associated graph theory data to obtain a trained electronic component detection model, wherein the electronic component detection model is a GCNs-LSTM model, and the GCNs-LSTM model includes two GCNs convolutional layers, two pooling layers, two input layers, two LSTM layers, and one output layer. The activation function of the GCNs convolutional layer is a ReLU function, and the GCNs-LSTM model removes the fully connected layer;
[0012] S5: Optimize the pooling method, pooling window size, regularization method, learning rate method and hyperparameters of the electronic component detection model based on the improved whale optimization algorithm;
[0013] S6: Collecting real-time data of electronic components, and using the trained electronic component detection model to perform fault detection on the real-time data to obtain a detection result.
[0014] Furthermore, each input layer is connected to two GCNs convolutional layers, and the two GCNs convolutional layers are divided into a first GCNs convolutional layer and a second GCNs convolutional layer;
[0015] The first GCNs convolutional layer and the second GCNs convolutional layer are respectively connected with a weight normalization layer, a ReLU nonlinear activation function layer, and a Sigmoid nonlinear activation function layer;
[0016] The first GCNs convolutional layer is used to process the single graph data;
[0017] The second GCNs convolutional layer is used to process the association graph data;
[0018] The single fault includes at least five states: short circuit, high resistance, open circuit, parameter drift, threshold voltage increase and vibration signal abnormality. The comprehensive fault includes the first type of comprehensive fault and the second type of comprehensive fault.
[0019] Furthermore, the preprocessing of the simulation data and the actual measured data to obtain feature data and an association relationship between the feature data, constructing single graph theory data based on the feature data, and constructing association graph theory data according to the association relationship of the feature data includes:
[0020] Preprocessing the simulation data and the actual measured data respectively to obtain processed first preprocessed feature data and second preprocessed feature data, wherein the preprocessing includes at least data normalization, clustering, and labeling, and the clustering is processed using a K-Means algorithm;
[0021] determining vertices of a single graph theory data based on the first preprocessed feature data and the second preprocessed feature data;
[0022] Determine adjacent vertices for each vertex based on the relationship between each feature data;
[0023] Connect each vertex with the adjacent vertex to form an edge, and determine the weight of the edge between each vertex and the adjacent vertex;
[0024] Obtaining single graph data based on vertices, adjacent vertices, edges, and edge weights, and obtaining associated graph data based on an association relationship between the first preprocessed feature data and the second preprocessed feature data;
[0025] Graph theory data are plotted using the Graphviz tool.
[0026] Furthermore, determining the vertices of the single graph data based on the first preprocessed feature data and the second preprocessed feature data includes:
[0027] Performing a first Fourier transform and a first Fourier feature extraction on the first preprocessed feature data and the second preprocessed feature data of each preset fault type to obtain a first signal feature of the first preprocessed feature data and a second signal feature of the second preprocessed feature data of each preset fault type;
[0028] Based on the first signal feature of the first pre-processed feature data and the second signal feature of the second pre-processed feature data of each preset fault type, a vertex of the single graph theory data is determined.
[0029] Furthermore, the electronic component detection model is trained based on the single graph theory data and the associated graph theory data to obtain a trained electronic component detection model, including:
[0030] Selecting a single graph theory data pair from the associated graph theory data and inputting the pair into the electronic component detection model to obtain a predicted detection result for each single graph theory data pair;
[0031] Calculating a loss function of the electronic component detection model based on the predicted detection result and the label of each single graph theory data pair in the associated graph theory data;
[0032] The model parameters are updated based on the loss function of the electronic component detection model until the model converges to obtain a trained electronic component detection model.
[0033] Furthermore, in said S1, obtaining simulation data of a preset fault type includes:
[0034] Acquire simulation data of each node of the electronic component under the single fault type;
[0035] For the single fault, determining a target fault node and a corresponding first fault type, and determining simulation data of the single fault based on simulation data of the target fault node under the first fault type;
[0036] For the comprehensive fault, determine that the electronic component is a second fault type or the row electronic component is a third fault type, and calculate the simulation data of the comprehensive fault based on the simulation data of the electronic component under the second fault type and the simulation data of the electronic component under the third fault type.
[0037] Furthermore, in S2, obtaining actual measured data of the preset fault type includes:
[0038] Set sensors at each node of electronic components in advance;
[0039] Using the sensor to collect actual measurement data of a preset number of groups of the electronic components under the preset fault type;
[0040] For any pair of preset fault types, the preset number of groups of actual measured data are preprocessed to obtain at least one group of processed actual measured data, and a weighted sum is calculated based on the at least one group of processed actual measured data as the actual measured data of the preset fault type, wherein the weight of each group of actual measured data is determined according to the position of the sensor.
[0041] Furthermore, in step S5, the pooling mode, pooling window size, regularization mode, learning rate mode and hyperparameters of the electronic component detection model optimized based on the improved whale optimization algorithm are:
[0042] The pooling method includes maximum pooling and average pooling;
[0043] The pooling window sizes include 2*2, 3*3 and 4*4;
[0044] The regularization methods include Dropout, L1 / L2 regularization and Batch-Normalization;
[0045] The learning rate mode includes exponential decay and step decay;
[0046] The hyperparameters include the number of neurons in the convolution layer, the size of the convolution kernel, and the number of neurons in the GCNs-LSTM layer.
[0047] Furthermore, in S5, the improved whale optimization algorithm includes: an initialization phase, a fitness evaluation phase, a search phase, an iterative search phase, and an optimal solution output phase;
[0048] The specific implementation process of the improved whale optimization algorithm is as follows:
[0049] D=|CX*(t)-X(t)|,
[0050] X(t+1)=X*(t)-AD,
[0051] t is the current iteration number, A and C are coefficients, X*(t) represents the best position so far, that is, the optimal solution, and X(t) represents the current search position;
[0052] A=2ar1-a,
[0053] C=2r2,
[0054] a=2-2t / Tmax, where r1 and r2 are random numbers in (0,1), the value of a decreases linearly from 2 to 0, t represents the current number of iterations, and Tmax represents the maximum number of iterations.
[0055] X(t+1)=X * (t)+D * *e bd *cos(2πd),
[0056] Among them, D * =|X * (t)-X(t)|, represents the distance between the position of the current parameter to be adjusted and the parameter to be adjusted, X * (t) is the best position vector so far, b is a constant used to define the shape of the spiral, and d is a random number in (-1,1).
[0057] D=|C*X rand -X(t)|,
[0058] X(t+1)=X rand -A*D,
[0059] Among them, X randIt is the position vector of the parameter to be adjusted that is randomly selected. The algorithm sets a search agent to be randomly selected when A≥1, sets the search boundary to the record of the adjustment parameter settings of the electronic component detection model for N months of operation, sets the number of iterations T, and finally searches for the optimal feature subset as the adjustment parameter set. The search conditions are the minimum number of iterations and the fastest convergence speed. The adjustment parameters include pooling method, pooling window size, regularization method, learning rate method and hyperparameters, and the records are saved in a preset database.
[0060] The beneficial effects of the electronic component fault detection method based on artificial intelligence of the present invention are as follows:
[0061] The present invention obtains simulation data of preset fault types; obtains actual measured data of the preset fault types; preprocesses the simulation data and the actual measured data to obtain feature data and the correlation relationship between the feature data, constructs single graph theory data based on the feature data, and constructs correlation graph theory data based on the correlation relationship of the feature data; constructs an electronic component detection model; optimizes the adjustment parameters of the electronic component detection model based on the improved whale optimization algorithm; collects real-time data of electronic components, and uses the trained electronic component detection model to perform fault detection on the real-time data to obtain detection results; the present invention optimizes various parameters of the deep learning network model, improves the detection speed and detection accuracy of the model, and the model outputs features with nonlinear characteristics, thereby improving the accuracy of detection of various fault states. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 The figure is a flow chart of an electronic component fault detection method based on artificial intelligence according to the present invention. DETAILED DESCRIPTION
[0063] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0064] Example 1
[0065] First, the technical terms involved in the embodiments of the present application are explained.
[0066] (1) GCNs model
[0067] Graph Convolutional Networks (GCNs) are a type of deep learning model specifically designed for processing graph-structured data. By extending convolution operations to graph structures, GCNs enable them to capture complex dependencies between nodes and effectively learn node representations.
[0068] Application scenarios:
[0069] Node classification: GCNs can be used to classify nodes based on graph structure and node features, such as fault classification and fault prediction of electronic components;
[0070] Node cluster detection: GCNs can help identify clusters of nodes with dense connections in the graph;
[0071] Network similarity: GCNs can be used to measure the similarity between two networks or sub-networks;
[0072] Unsupervised learning: GCNs can also be used for unsupervised learning tasks, such as learning low-dimensional representations of nodes through autoencoders;
[0073] Text classification and information extraction: GCNs can use words and documents in text as nodes and build edges based on co-occurrence information to perform text classification or information extraction;
[0074] GCNs have become an important tool for processing graph-structured data due to their excellent performance and flexibility. With the deepening of research, GCNs and their variants have shown great potential in processing more complex graph-structured data.
[0075] Figure 1 The figure is a flow chart of an electronic component fault detection method based on artificial intelligence according to the present invention.
[0076] The method comprises steps S1-S6:
[0077] S1: Acquire simulation data of preset fault types, where the preset fault types are divided into single fault and combined fault;
[0078] S2: Obtaining actual measured data of the preset fault type;
[0079] S3: Preprocessing the simulation data and the actual measured data to obtain feature data and correlation relationships between the feature data, constructing single graph data based on the feature data, and constructing correlation graph data based on the correlation relationships between the feature data;
[0080] S4: constructing an electronic component detection model, and training the electronic component detection model based on the single graph theory data and the associated graph theory data to obtain a trained electronic component detection model, wherein the electronic component detection model is a GCNs-LSTM model, and the GCNs-LSTM model includes two GCNs convolutional layers, two pooling layers, two input layers, two LSTM layers, and one output layer. The activation function of the GCNs convolutional layer is a ReLU function, and the GCNs-LSTM model removes the fully connected layer;
[0081] S5: Optimize the pooling method, pooling window size, regularization method, learning rate method and hyperparameters of the electronic component detection model based on the improved whale optimization algorithm;
[0082] S6: Collecting real-time data of electronic components, and using the trained electronic component detection model to perform fault detection on the real-time data to obtain a detection result.
[0083] Specifically, the preset fault types are divided into single faults and combined faults. Table 1 illustrates various fault conditions of typical electronic components. It may be that all nodes of the electronic components are normal, some nodes of the electronic components are normal and some nodes are faulty, or all nodes are faulty.
[0084] Table 1 Various fault conditions under two fault states: single fault and combined fault:
[0085]
[0086]
[0087] As shown in Table 1, the single fault includes at least five states: short circuit, high resistance, open circuit, parameter drift, threshold voltage increase, and abnormal vibration signal. Typical electronic components include diodes, triodes, transistors, MOS tubes, resistors, and capacitors. The single fault means that only one node of the electronic component fails.
[0088] As shown in Table 1, the comprehensive fault refers to different faults occurring at multiple nodes of a typical electronic component. The comprehensive fault includes a first type of comprehensive fault and a second type of comprehensive fault. The first type of comprehensive fault refers to the same fault state occurring at each node of the electronic component, such as all being in a high-impedance state. The second type of comprehensive fault refers to different fault states occurring at each node of the electronic component, such as one node being in a high-impedance state and another being in an open circuit state.
[0089] Furthermore, each input layer is connected to two GCNs convolutional layers, and the two GCNs convolutional layers are divided into a first GCNs convolutional layer and a second GCNs convolutional layer;
[0090] The first GCNs convolutional layer and the second GCNs convolutional layer are respectively connected with a weight normalization layer, a ReLU nonlinear activation function layer, and a Sigmoid nonlinear activation function layer;
[0091] The first GCNs convolutional layer is used to process the single graph data;
[0092] The second GCNs convolutional layer is used to process the association graph data;
[0093] The single fault includes at least five states: short circuit, high resistance, open circuit, parameter drift, threshold voltage increase and vibration signal abnormality. The comprehensive fault includes the first type of comprehensive fault and the second type of comprehensive fault.
[0094] Furthermore, the preprocessing of the simulation data and the actual measured data to obtain feature data and an association relationship between the feature data, constructing single graph theory data based on the feature data, and constructing association graph theory data according to the association relationship of the feature data includes:
[0095] Preprocessing the simulation data and the actual measured data respectively to obtain processed first preprocessed feature data and second preprocessed feature data, wherein the preprocessing includes at least data normalization, clustering, and labeling, and the clustering is processed using a K-Means algorithm;
[0096] determining vertices of a single graph theory data based on the first preprocessed feature data and the second preprocessed feature data;
[0097] Determine adjacent vertices for each vertex based on the relationship between each feature data;
[0098] Connect each vertex with the adjacent vertex to form an edge, and determine the weight of the edge between each vertex and the adjacent vertex;
[0099] Obtaining single graph data based on vertices, adjacent vertices, edges, and edge weights, and obtaining associated graph data based on an association relationship between the first preprocessed feature data and the second preprocessed feature data;
[0100] Graph theory data is drawn using the Graphviz tool;
[0101] Specifically, in GCNs, a graph consists of vertices and edges, usually represented as G = (V, E), where V is a set of vertices, E is a set of edges, each vertex has a feature vector, and the edge represents the connection relationship between nodes. The adjacency matrix A of the graph represents the connection between vertices, where A ij Indicates whether there is an edge connection between vertex i and vertex j;
[0102] To avoid the gradient vanishing or gradient exploding problem, the GCNs in this embodiment normalizes the adjacency matrix and uses the square root of the degree matrix D to normalize A, obtaining
[0103] GCNs convolve the features of the vertex and aggregate the features of the vertex with the features of its neighbors, calculated by the following formula:
[0104]
[0105] Among them H n is the vertex feature matrix of the nth layer, W n is a learnable weight matrix and Sigmoid is a nonlinear activation function.
[0106] Furthermore, determining the vertices of the single graph data based on the first preprocessed feature data and the second preprocessed feature data includes:
[0107] Performing a first Fourier transform and a first Fourier feature extraction on the first preprocessed feature data and the second preprocessed feature data of each preset fault type to obtain a first signal feature of the first preprocessed feature data and a second signal feature of the second preprocessed feature data of each preset fault type;
[0108] Based on the first signal feature of the first pre-processed feature data and the second signal feature of the second pre-processed feature data of each preset fault type, a vertex of the single graph theory data is determined.
[0109] Furthermore, the electronic component detection model is trained based on the single graph theory data and the associated graph theory data to obtain a trained electronic component detection model, including:
[0110] Selecting a single graph theory data pair from the associated graph theory data and inputting the pair into the electronic component detection model to obtain a predicted detection result for each single graph theory data pair;
[0111] Calculating a loss function of the electronic component detection model based on the predicted detection result and the label of each single graph theory data pair in the associated graph theory data;
[0112] The model parameters are updated based on the loss function of the electronic component detection model until the model converges to obtain a trained electronic component detection model.
[0113] Furthermore, in said S1, obtaining simulation data of a preset fault type includes:
[0114] Acquire simulation data of each node of the electronic component under the single fault type;
[0115] For the single fault, determining a target fault node and a corresponding first fault type, and determining simulation data of the single fault based on simulation data of the target fault node under the first fault type;
[0116] For the comprehensive fault, determine that the electronic component is a second fault type or the row electronic component is a third fault type, and calculate the simulation data of the comprehensive fault based on the simulation data of the electronic component under the second fault type and the simulation data of the electronic component under the third fault type.
[0117] Furthermore, in S2, obtaining actual measured data of the preset fault type includes:
[0118] Set sensors at each node of electronic components in advance;
[0119] Using the sensor to collect actual measurement data of a preset number of groups of the electronic components under the preset fault type;
[0120] For any pair of preset fault types, the preset number of groups of actual measured data are preprocessed to obtain at least one group of processed actual measured data, and a weighted sum is calculated based on the at least one group of processed actual measured data as the actual measured data of the preset fault type, wherein the weight of each group of actual measured data is determined according to the position of the sensor.
[0121] Furthermore, in step S5, the pooling mode, pooling window size, regularization mode, learning rate mode and hyperparameters of the electronic component detection model optimized based on the improved whale optimization algorithm are:
[0122] The pooling method includes maximum pooling and average pooling;
[0123] The pooling window sizes include 2*2, 3*3 and 4*4;
[0124] The regularization methods include Dropout, L1 / L2 regularization and Batch-Normalization;
[0125] The learning rate mode includes exponential decay and step decay;
[0126] The hyperparameters include the number of neurons in the convolution layer, the size of the convolution kernel, and the number of neurons in the GCNs-LSTM layer.
[0127] Furthermore, in S5, the improved whale optimization algorithm includes: an initialization phase, a fitness evaluation phase, a search phase, an iterative search phase, and an optimal solution output phase;
[0128] The specific implementation process of the improved whale optimization algorithm is as follows:
[0129] D=|CX*(t)-X(t)|,
[0130] X(t+1)=X*(t)-AD,
[0131] t is the current iteration number, A and C are coefficients, X*(t) represents the best position so far, that is, the optimal solution, and X(t) represents the current search position;
[0132] A=2ar1-a,
[0133] C=2r2,
[0134] a=2-2t / Tmax, where r1 and r2 are random numbers in (0,1), the value of a decreases linearly from 2 to 0, t represents the current number of iterations, and Tmax represents the maximum number of iterations.
[0135] X(t+1)=X * (t)+D * *e bd *cos(2πd),
[0136] Among them, D * =|X * (t)-X(t)|, represents the distance between the position of the current parameter to be adjusted and the parameter to be adjusted, X * (t) is the best position vector so far, b is a constant used to define the shape of the spiral, and d is a random number in (-1,1).
[0137] D=|C*X rand -X(t)|,
[0138] X(t+1)=X rand -A*D,
[0139] Among them, X rand It is the position vector of the parameter to be adjusted that is randomly selected. The algorithm sets a search agent to be randomly selected when A≥1, sets the search boundary to the record of the adjustment parameter settings of the electronic component detection model for N months of operation, sets the number of iterations T, and finally searches for the optimal feature subset as the adjustment parameter set. The search conditions are the minimum number of iterations and the fastest convergence speed. The adjustment parameters include pooling method, pooling window size, regularization method, learning rate method and hyperparameters, and the records are saved in a preset database.
[0140] The beneficial effects of the electronic component fault detection method based on artificial intelligence of the present invention are as follows:
[0141] The present invention obtains simulation data of preset fault types; obtains actual measured data of the preset fault types; preprocesses the simulation data and the actual measured data to obtain feature data and the correlation relationship between the feature data, constructs single graph theory data based on the feature data, and constructs correlation graph theory data based on the correlation relationship of the feature data; constructs an electronic component detection model; optimizes the adjustment parameters of the electronic component detection model based on the improved whale optimization algorithm; collects real-time data of electronic components, and uses the trained electronic component detection model to perform fault detection on the real-time data to obtain detection results; the present invention optimizes various parameters of the deep learning network model, improves the detection speed and detection accuracy of the model, and the model outputs features with nonlinear characteristics, thereby improving the accuracy of detection of various fault states.
[0142] The present invention and its embodiments are described above. This description is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual content is not limited to this. In short, if ordinary technicians in this field are inspired by it and do not depart from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.
Claims
1. An electronic component fault detection method based on artificial intelligence, characterized in that: The method comprises: S1: Acquire simulation data of preset fault types, where the preset fault types are divided into single fault and combined fault; S2: Acquire actual detection data of the preset fault type; S3: preprocessing the simulation data and the actual measured data to obtain feature data and the correlation relationship between the feature data, constructing single graph data based on the feature data, and constructing correlation graph data according to the correlation relationship between the feature data; S4: constructing an electronic component detection model, and training the electronic component detection model based on the single graph theory data and the associated graph theory data to obtain a trained electronic component detection model, wherein the electronic component detection model is a GCNs-LSTM, and the GCNs-LSTM model includes 2 GCNs convolutional layers, 2 pooling layers, two input layers, two LSTM layers, and one output layer, and the activation function of the GCNs convolutional layer is a ReLU function, wherein the GCNs-LSTM model removes the fully connected layer; S5: Optimize the pooling method, pooling window size, regularization method, learning rate method and hyperparameters of the electronic component detection model based on the improved whale optimization algorithm; S6: Collecting real-time data of electronic components, and using the trained electronic component detection model to perform fault detection on the real-time data to obtain detection results.
2. The method for detecting electronic component faults based on artificial intelligence according to claim 1, characterized in that: Each input layer is connected to two GCNs convolutional layers, and the two GCNs convolutional layers are divided into a first GCNs convolutional layer and a second GCNs convolutional layer; The first GCNs convolution layer and the second GCNs convolution layer are connected to a weight normalization layer, a ReLU nonlinear activation function layer, and a Sigmoid nonlinear activation function layer respectively; The first GCNs convolutional layer is used to process the single graph theory data; The second GCNs convolutional layer is used to process the association graph data; The single fault includes at least five states: short circuit, high resistance, open circuit, parameter drift, threshold voltage increase and abnormal vibration signal. The comprehensive fault includes a first type of comprehensive fault and a second type of comprehensive fault.
3. The method for detecting electronic component faults based on artificial intelligence according to claim 2, characterized in that: The preprocessing of the simulation data and the actual measured data to obtain feature data and the correlation relationship between the feature data, constructing single graph data based on the feature data, and constructing correlation graph data according to the correlation relationship between the feature data, includes: Preprocessing the simulation data and the actual measured data respectively to obtain processed first preprocessed feature data and second preprocessed feature data, wherein the preprocessing at least includes data normalization, clustering, and labeling, and the clustering is processed using a K-Means algorithm; Determining vertices of a single graph theory data based on the first preprocessed feature data and the second preprocessed feature data; Determine adjacent vertices for each vertex based on the relationship between each feature data; Connect each vertex with the adjacent vertex to form an edge, and determine the weight of the edge between each vertex and the adjacent vertex; Obtaining single graph data based on vertices, adjacent vertices, edges, and edge weights, and obtaining associated graph data based on an association relationship between the first preprocessed feature data and the second preprocessed feature data; Graph data are plotted using the Graphviz tool.
4. The method for detecting electronic component faults based on artificial intelligence according to claim 3, characterized in that: The step of determining the vertices of a single graph data based on the first preprocessed feature data and the second preprocessed feature data comprises: Performing a first Fourier transform and a first Fourier feature extraction on the first preprocessed feature data and the second preprocessed feature data of each preset fault type to obtain a first signal feature of the first preprocessed feature data and a second signal feature of the second preprocessed feature data of each preset fault type; Based on the first signal feature of the first preprocessing feature data and the second signal feature of the second preprocessing feature data of each preset fault type, a vertex of the single graph theory data is determined.
5. The method for detecting electronic component faults based on artificial intelligence according to claim 4, characterized in that: The step of training the electronic component detection model based on the single graph theory data and the associated graph theory data to obtain the trained electronic component detection model includes: Selecting a single graph theory data pair from the associated graph theory data and inputting it into the electronic component detection model to obtain a predicted detection result of each single graph theory data pair; Calculating a loss function of the electronic component detection model based on the predicted detection result and the label of each single graph theory data pair in the associated graph theory data; Based on the loss function of the electronic component detection model, the model parameters are updated until the model converges to obtain a trained electronic component detection model.
6. The method for detecting electronic component faults based on artificial intelligence according to claim 5, characterized in that: In S1, obtaining simulation data of a preset fault type includes: Acquire simulation data of each node of the electronic component under the single fault type; For the single fault, determine a target fault node and a corresponding first fault type, and determine simulation data of the single fault based on simulation data of the target fault node in the first fault type; For the comprehensive fault, determine that the electronic component is a second fault type or the row of electronic components is a third fault type, and calculate the simulation data of the comprehensive fault based on the simulation data of the electronic component under the second fault type and the simulation data of the electronic component under the third fault type.
7. The method for detecting electronic component faults based on artificial intelligence as claimed in claim 6, characterized in that: In S2, obtaining actual measured data of the preset fault type includes: Pre-set sensors at each node of electronic components; Using the sensor to collect actual measurement data of a preset number of groups of the electronic components under the preset fault type; For any pair of preset fault types, the preset number of groups of actual measured data are preprocessed to obtain at least one group of processed actual measured data, and a weighted sum is calculated based on the at least one group of processed actual measured data as the actual measured data of the preset fault type, wherein the weight of each group of actual measured data is determined according to the position of the sensor.
8. The method for detecting electronic component faults based on artificial intelligence as claimed in claim 7, characterized in that: In step S5, the pooling method, pooling window size, regularization method, learning rate method and hyperparameters of the electronic component detection model optimized based on the improved whale optimization algorithm are: The pooling method includes maximum pooling and average pooling; The pooling window sizes include 2*2, 3*3 and 4*4; The regularization methods include Dropout, L1 / L2 regularization and Batch-Normalization; The learning rate mode includes exponential decay and step decay; The hyperparameters include the number of neurons in the convolution layer, the size of the convolution kernel, and the number of neurons in the GCNs-LSTM layer.
9. The method for detecting electronic component faults based on artificial intelligence as claimed in claim 8, characterized in that: In S5, the improved whale optimization algorithm includes: an initialization phase, a fitness evaluation phase, a search phase, an iterative search phase, and an optimal solution output phase; The specific implementation process of the improved whale optimization algorithm is as follows: D=|CX*(t)-X(t)|, X(t+1)=X*(t)-AD, t is the current iteration number, A and C are the coefficients, X*(t) represents the best position so far, that is, the optimal solution, and X(t) represents the current search position; A=2ar1-a, C＝2r2， a=2-2t / Tmax, Among them, r1 and r2 are random numbers in (0,1), the value of a decreases linearly from 2 to 0, t represents the current number of iterations, and Tmax represents the maximum number of iterations. X(t+1)=X * (t)+D * *e bd *cos(2πd), Among them, D * =|X * (t)-X(t)|, represents the distance between the current position of the parameter to be adjusted and the parameter to be adjusted, X * (t) is the best position vector so far, b is a constant used to define the shape of the spiral, and d is a random number in (-1,1). D=∣C*X rand -X(t)∣, X(t+1)=X rand -A*D, Among them, X rand It is the position vector of the parameter to be adjusted selected randomly. The algorithm sets that when A≥1, a search agent is randomly selected, and the search boundary is set as the record of the adjustment parameter settings of the electronic component detection model running for N months. The number of iterations is set to T. Finally, the optimal feature subset obtained by the search is the adjustment parameter set. The search conditions are the minimum number of iterations and the fastest convergence speed. The adjustment parameters include pooling method, pooling window size, regularization method, learning rate method and hyperparameters. The records are saved in a preset database.
Citation Information
Patent Citations
Electronic component fault diagnosis system based on artificial intelligence
CN117830309A