Electronic component detection data processing method and system

Abstract

The invention relates to the technical field of electronic component detection, and discloses an electronic component detection data processing method and system.The electronic component detection data processing method comprises the steps that detection data of multiple electronic components are collected and preprocessed, and high-quality component detection data and system topological structure information are obtained; calculating multi-dimensional similarity among a plurality of electronic components, and establishing a component similarity network; performing multi-scale time sequence feature extraction, identifying an abnormal mode and calculating an abnormal degree score of the component; constructing a time sequence heterogeneous graph network; according to the time sequence heterogeneous graph network, applying a graph attention network model to predict a propagation path of the electronic component fault; electronic component detection is expanded from traditional single-point analysis to system-level analysis, a propagation path of a fault in a system can be predicted, and potential system fault risks can be warned in advance.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic component detection, and more specifically, it relates to a method and system for processing electronic component detection data. Background Art

[0002] With the continuous improvement of the complexity of electronic systems, the number of components in integrated circuits and electronic devices has increased exponentially, and the interaction and dependency relationships among components have become increasingly complex. In practical applications, the performance degradation or failure of a single component often affects other components through circuit connection relationships, forming complex fault propagation paths, which pose serious potential risks to system stability.

[0003] In the prior art, there are mainly the following deficiencies in the detection of electronic components: First, traditional detection methods usually analyze the performance parameters of each component independently, compare them with preset thresholds to determine whether the current state is qualified, and ignore the correlation information between components and the impact of potential cascading failures; Second, conventional detection only focuses on the current state of components, and cannot mine the performance degradation trend from historical detection data, lacking the ability of forward-looking warning; Third, the prior art lacks the ability to analyze the fault propagation paths in large-scale integrated circuits and electronic systems and cannot predict the system-level fault risks.

[0004] Therefore, there is an urgent need for a technical solution that can integrate the detection data of multiple components and system topology information to achieve the prediction of the propagation of group faults of electronic components, so as to early warn of potential system fault risks and ensure the stable and reliable operation of electronic systems. Summary of the Invention

[0005] The present invention provides a method and system for processing electronic component detection data, which solves the technical problems in the related art of ignoring the correlation information between components, lacking the ability of forward-looking warning, and being unable to predict system-level fault risks.

[0006] The present invention provides a method for processing electronic component detection data, including: Collecting the detection data of multiple electronic components and performing preprocessing to obtain high-quality component detection data and system topology structure information; Calculating the multi-dimensional similarity between multiple electronic components based on the component detection data to establish a component similarity network; Performing multi-scale time-series feature extraction based on the component detection data, identifying abnormal patterns and calculating the component abnormality degree score; Based on the component similarity network, system topology structure information and component abnormality degree score, constructing a time-series heterogeneous graph network; According to the time-series heterogeneous graph network, applying a graph attention network model to predict the propagation path of electronic component faults.

[0007] Furthermore, the detection data includes static parameter data, electrical parameter data, thermal parameter data, and timing data.

[0008] Furthermore, the detection data includes static parameter data, electrical parameter data, thermal parameter data, and timing data.

[0009] Calculating the multi-dimensional similarity between multiple electronic components includes: Extracting the static attributes and dynamic performance characteristics of components to construct feature vectors; Calculating the similarity between components using cosine similarity; Setting a similarity threshold and nearest neighbor constraint to construct a component similarity network.

[0010] Furthermore, the detection data includes static parameter data, electrical parameter data, thermal parameter data, and timing data.

[0011] The multi-scale timing feature extraction includes: Applying wavelet transform to the timing data for multi-scale decomposition; Extracting statistical features from the decomposition results, including mean, variance, skewness, kurtosis, and entropy; Constructing an abnormal pattern library and using the dynamic time warping algorithm to calculate the distance between the component to be measured and the abnormal pattern; Generating a component abnormality degree score.

[0012] Furthermore, the detection data includes static parameter data, electrical parameter data, thermal parameter data, and timing data.

[0013] The timing heterogeneous graph network includes: Representing electronic components as nodes in a graph network; Constructing three types of relationship edges, including physical connection relationship, functional dependency relationship, and similarity relationship; Integrating the timing information into the graph network to form a timing heterogeneous graph network.

[0014] Furthermore, the detection data includes static parameter data, electrical parameter data, thermal parameter data, and timing data.

[0015] The graph attention network model includes: An input layer that receives node feature and edge relationship information; Multiple graph attention layers, each layer containing multiple attention heads for learning the influence weights between nodes; A heterogeneous relationship fusion mechanism that integrates the attention coefficients of different types of edges; A timing information integration mechanism that captures the time dynamic characteristics of fault propagation; The output layer generates the influence transfer probability between components.

[0016] Further, the detection data includes static parameter data, electrical parameter data, thermal parameter data, and timing data.

[0017] The heterogeneous relationship fusion mechanism is achieved by assigning different attention coefficients to different types of relationship edges and performing weighted fusion.

[0018] Further, the detection data includes static parameter data, electrical parameter data, thermal parameter data, and timing data.

[0019] The timing information integration mechanism includes: Constructing a timing window with a time step of ; Calculating the attention weights at each time step; Applying the timing attention mechanism to integrate information from different time steps.

[0020] Further, the detection data includes static parameter data, electrical parameter data, thermal parameter data, and timing data.

[0021] The predicted propagation path of electronic component failures includes: Applying the graph attention network model to analyze the influence transfer probability between nodes; Monte Carlo tree search simulation, based on the calculated influence transfer probability between nodes, applying the Monte Carlo tree search method to simulate the propagation process of failures in the system; Identifying critical components and propagation paths in the system based on the simulation results.

[0022] An electronic component detection data processing system for executing the above-mentioned electronic component detection data processing method includes: A data preprocessing module for collecting and preprocessing the detection data of multiple electronic components; A similarity analysis module for calculating the multi-dimensional similarity between multiple electronic components and establishing a component similarity network; An anomaly detection module for extracting multi-scale timing features from the component detection data, identifying anomaly patterns, and calculating the component anomaly degree score; A graph network construction module for constructing a timing heterogeneous graph network; A failure propagation prediction module for applying the graph attention network model to predict the propagation path of electronic component failures.

[0023] The beneficial effects of the present invention are as follows: It expands the detection of electronic components from traditional single-point analysis to system-level analysis, can predict the propagation path of faults in the system, and early warns of potential system fault risks. By constructing a temporal heterogeneous graph network model, it captures the physical connections, functional dependencies, and similarity relationships between components, revealing the deep mechanism of fault propagation; Using a multi-dimensional component similarity calculation method, it fully explores the correlation of component group data, improves the detection accuracy, is applicable to large-scale component batch detection scenarios, and can quickly identify the potential risks of similar components based on existing fault cases.

[0024] Through multi-scale temporal feature extraction and anomaly pattern recognition technologies, it realizes the early identification of the performance degradation trend of electronic components, can capture weak anomaly signals that cannot be detected by traditional threshold detection methods, improves the sensitivity and specificity of prediction, and reduces the false alarm and missed alarm rates.

[0025] The fault propagation path prediction method based on graph attention network and Monte Carlo tree search can automatically identify key components and high-risk propagation paths in the system. By strengthening the monitoring and maintenance of these key points, the failure rate of the system caused by early component failures is reduced.

[0026] Compared with traditional independent detection methods, this method can discover common problems between similar components, can predict potential risks at the initial stage of new batch component production, reduces the problem discovery time, and improves production efficiency and product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flowchart of a method for processing electronic component detection data in the present invention; Figure 2 is a flowchart of step 1 of the present invention; Figure 3 is a flowchart of step 2 of the present invention; Figure 4 is a flowchart of step 3 of the present invention; Figure 5 is a flowchart of step 4 of the present invention; Figure 6 is a flowchart of step 5 of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0028] Reference will now be made to exemplary embodiments to discuss the subject matter described herein. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and that changes may be made to the functions and arrangements of the elements discussed without departing from the scope of protection of the content of this specification. Each example may omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples may be combined in other examples.

[0029] In at least one embodiment of the present invention, a method for processing electronic component detection data is disclosed, as Figures 1 to 6 shown, including the following steps: Step 1, collect the detection data of multiple electronic components and perform preprocessing to obtain high-quality component detection data and system topology structure information; This step is used to obtain the historical detection data of electronic components and perform preprocessing to provide high-quality input data for subsequent analysis. Specifically, it includes the following sub-steps: Step 1.1, multi-dimensional detection data collection; Collect the historical detection data of electronic components, including multi-dimensional time series data of electrical parameters (such as voltage, current, resistance), thermal parameters (such as temperature, thermal resistance), acoustic parameters (such as vibration frequency, amplitude), and other performance indicators; the collection frequency is configured according to the component characteristics and application scenarios, and the collection frequency of key parameters is usually set to 5 - 10 times per second, and the non-key parameters can be set to 1 - 5 times per minute; at the same time, record the basic information of the components, including metadata such as manufacturing batch, usage environment, and operation duration.

[0030] Step 1.2, data preprocessing; Perform preprocessing on the collected raw data, mainly including: Denoising processing: Apply the wavelet denoising algorithm to remove the random noise in the data and retain the effective signal. Specifically, perform wavelet transform on the signal to obtain wavelet coefficients at different scales , and its calculation formula is: ; where is the wavelet coefficient, represents the wavelet basis function, is the scale parameter, is the translation parameter, represents the complex conjugate of is the time variable, represents the integration over all time variables; Then perform soft threshold processing on the wavelet coefficients, and the threshold can be determined by the following formula: ; where is the threshold value, is the estimated value of the noise standard deviation, is the number of data points, represents the logarithmic operation; Finally, the denoised signal is reconstructed through the inverse wavelet transform.

[0031] Data standardization: Normalize different types of parameter data so that their value ranges are unified within the range of [0, 1], which is convenient for subsequent analysis; For the parameter , its standardized value is calculated as follows: ; where and are the original data and the standardized data respectively, and are the minimum value and the maximum value of this parameter respectively.

[0032] Time alignment: For parameter data with different acquisition frequencies, perform time alignment through interpolation and resampling methods to ensure that all parameters have corresponding data values at the same time point; For low sampling rate data, use the cubic spline interpolation method for filling.

[0033] Step 1.3, construction of system topology structure information; According to the circuit schematic diagram and the system design document, extract the connection relationship between electronic components and construct the system topology structure information; Specifically include: Component node representation: Represent each electronic component as a node and record its type, function, location and other attribute information; Connection relationship representation: Extract the physical connection relationship between components, including direct connection and indirect connection, to form a connection matrix , where represents that there is a direct connection between component and component , represents that there is no direct connection; Functional dependency relationship representation: Analyze the functional dependency relationship between components and construct a functional dependency matrix , where represents the degree of functional dependency of component on component , and the value range is [0, 1]; After the processing of this step, high-quality component detection data and system topology structure information are obtained, providing a basis for subsequent analysis.

[0034] Step 2: Calculate the multi-dimensional similarity between multiple electronic components based on the component detection data, and establish a component similarity network. In this step, based on the preprocessed component detection data obtained in Step 1, the multi-dimensional similarity between components is calculated, and a component similarity network is constructed, providing a basis for subsequent collaborative fault prediction. The standardized detection data output by Step 1 enables effective comparison of the performance parameters between different components, thereby accurately evaluating their similarity. This step includes the following sub-steps: Step 2.1: Construction of multi-dimensional feature vectors For each electronic component, construct multi-dimensional feature vectors, including the following types of features: Static features: Include inherent attribute information such as the manufacturing batch, material type, specification parameters, and service life of the component. Environmental features: Include environmental condition parameters such as the operating temperature, humidity, and vibration intensity of the component. Performance features: Include statistical features of the key performance indicators of the component within the normal operating range, such as mean, variance, kurtosis, skewness, etc. Temporal features: Extract the features of the component performance parameters changing with time, including trend features, periodic features, and abnormal fluctuation features, etc.

[0035] Integrate the above features to form the feature vector of the component , where represents the index of the component; in order to balance the importance of different types of features, weight coefficients are assigned to each type of feature to form a weighted feature vector.

[0036] Step 2.2: Calculation of the similarity matrix Based on the feature vectors of the components, calculate the similarity between any two components and construct a similarity matrix ; Considering the differences of different types of features, multiple measurement methods are used to calculate the similarity: For numerical features, the cosine similarity calculation method is adopted: ; where represents the cosine similarity between component and component , and respectively represent the feature vectors of component and component , represents the vector Euclidean norm; For the case considering the feature distribution, the Mahalanobis distance is used to calculate the similarity (the Mahalanobis distance takes into account the feature distribution and can more accurately measure the distance between components in the feature space under different scales and correlations): ; where represents the similarity between component and component based on the Mahalanobis distance, represents the inverse matrix of the feature covariance matrix, which is used to consider the correlation between features and the variances of different features, represents the vector difference in its transposed form, represents the transpose operator.

[0037] For categorical features, the Jaccard similarity coefficient is used: ; where represents the Jaccard similarity coefficient between component and component , and represent the union and intersection operations of sets respectively, represents the number of elements in the set.

[0038] The above different types of similarities are weighted and fused to obtain the comprehensive similarity : ; where represents the comprehensive similarity between component and component , , and represent the weight coefficients of the cosine similarity, Mahalanobis distance similarity, and Jaccard similarity coefficient respectively.

[0039] Step 2.3, Component similarity network construction; Based on the calculated similarity matrix , a component similarity network is constructed: ; where represents the component similarity network, represents the set of nodes (each node represents a component), represents the set of edges (indicating the similarity relationship connections between components), Represents the set of edge weights (the weight of each edge corresponds to the similarity value in the similarity matrix).

[0040] To reduce the network complexity, a similarity threshold is set , and only when the similarity between components is greater than or equal to the threshold, a connection edge is established between the corresponding nodes, and the weight of the edge is set to . The threshold can be adjusted according to the specific application scenario and the number of components, and the general value range is [0.7, 0.9].

[0041] Meanwhile, to more accurately describe the similarity relationship between components, -nearest neighbor ( -NN) constraint is introduced. For each component node, keep the connections with the neighbor nodes with the highest similarity to it, where is usually determined according to the square root of the total number of components.

[0042] Through this step, a network structure describing the similarity relationship between components is established, providing key support for subsequent collaborative fault prediction.

[0043] Step 3: Based on the component detection data, perform multi-scale time series feature extraction, identify abnormal patterns, and calculate the abnormal degree score of components; In this step, the time series detection data of components obtained in Step 1 is used for multi-scale feature extraction, identifying abnormal patterns and calculating the abnormal degree score of components; by deeply analyzing the high-quality time series data output in Step 1, the characteristics and laws of the change of component performance parameters over time can be mined, providing a basis for abnormal state detection. This step includes the following sub-steps: Step 3.1: Multi-scale time series feature extraction; Perform multi-scale time series feature extraction on the performance data of electronic components to capture signal features at different time scales: Wavelet multi-scale decomposition: Apply wavelet transform to perform multi-scale decomposition on the time series data of component performance, capturing both short-term fluctuations and long-term trends at the same time. For the time series signal , select a wavelet basis function suitable for the characteristics of components (such as Daubechies wavelet, Symlet wavelet, etc.), and perform -layer wavelet decomposition to obtain a set of approximation coefficients and several sets of detail coefficients, respectively reflecting the characteristics of the signal at different frequency scales: ; where , , represent the first layer, the second layer, the The detail coefficient of the layer, indicating the number of layers of wavelet decomposition; For the detection data of electronic components, usually select to the decomposition level.

[0044] Taking the leakage current parameter of a capacitor as an example, after wavelet decomposition of the leakage current time series data of a certain capacitor under normal working conditions, the low-frequency stable trend signal and the fluctuation signals of different frequency bands can be separated. In practical applications, the low-frequency trend component usually reflects the slow deterioration process of the capacitor performance, while the medium-high frequency fluctuation component may capture sudden abnormal events.

[0045] Time series statistical feature extraction: Extract statistical features for the decomposition coefficients of different scales, including mean, variance, skewness, kurtosis, entropy, etc., to form a time series statistical feature vector: ; where represents the time series statistical feature vector, represents the mean of the approximation coefficient of the th layer, reflecting the overall level of the signal; represents the variance of the approximation coefficient of the th layer, reflecting the degree of signal fluctuation; represents the skewness of the approximation coefficient of the th layer, reflecting the asymmetry of signal distribution; represents the kurtosis of the approximation coefficient of the th layer, reflecting the sharpness of signal distribution; represents the entropy value of the approximation coefficient of the th layer, reflecting the complexity and uncertainty of the signal; represents the mean of the detail coefficient of the first layer, represents the entropy value of the detail coefficient of the th layer, capturing the signal characteristics at different frequency scales.

[0046] Trend feature extraction: Apply the EMD (Empirical Mode Decomposition) method or the SSA (Singular Spectrum Analysis) method to extract the trend component of the time series data, and calculate features such as trend slope and curvature to form a trend feature vector ; Periodic feature extraction: Based on FFT (Fast Fourier Transform) or autocorrelation analysis, identify the periodic patterns in the time series data, and extract features such as the main frequency component, amplitude, and phase to form a periodic feature vector ; Integrate the above various features to form a comprehensive time series feature vector: ; where Represents the comprehensive timing feature vector, which includes the timing statistical feature vector , trend feature vector and periodic feature vector , comprehensively describes the timing characteristics of the performance parameters of components, and serves as the feature representation of the performance degradation state of components.

[0047] Step 3.2, construction of the abnormal mode library; Construct the abnormal mode feature library of electronic components and store the feature sequences of known failure cases: Collect known failure cases: Collect the confirmed component failure cases from historical data, and record the complete timing data of their performance parameters, especially the data change trend before failure; Extract failure mode features: Apply the above multi-scale feature extraction method to the timing data of each failure case to obtain the feature vectors in different time windows before failure, and form the failure mode feature sequence; Abnormal mode clustering: Use the density clustering algorithm (such as the DBSCAN algorithm) to perform clustering analysis on the feature sequences of all failure cases to identify typical failure mode categories. For the clustering results, calculate the center point and standard deviation of each category as the feature representation of this type of abnormal mode; Abnormal mode library storage: Save the abnormal mode features obtained by clustering and their corresponding failure types, severity, development speed and other information into the abnormal mode feature library, providing a reference benchmark for subsequent abnormal detection.

[0048] Step 3.3, scoring the abnormal degree of components; Based on the multi-scale timing features and the abnormal mode library, calculate the abnormal degree score of the component to be tested: Application of the dynamic time warping algorithm: For the timing feature sequence of the component to be tested and the reference pattern in the abnormal mode library , apply the dynamic time warping (DTW) algorithm to calculate the minimum deformation distance between them; where represents the dynamic time warping distance between the timing feature sequence of the component to be tested and the reference pattern in the abnormal mode library, represents the time alignment function, which is used to optimally match the timing features of different lengths and rates, represents the index of the reference feature point corresponding to the th test feature point, realizing the elastic alignment of the timing data in the time dimension; Represents the distance metric function between feature points, usually using Euclidean distance or Mahalanobis distance; Represents the operation of finding the minimum value; Represents the summation of the distances of all test feature points, Represents the length of the test feature sequence.

[0049] The DTW algorithm can effectively handle problems such as stretching, offset, and different sampling rates of time series data on the time axis, and is particularly suitable for the situation where the performance degradation rates of electronic components are inconsistent under different working conditions. Through the dynamic programming method, the DTW algorithm can find the best matching path between two time series, even if these series differ in time length, sampling frequency, or degradation speed. This feature enables the system to accurately identify the abnormal patterns of components, even if the time course of the abnormal development does not exactly match the reference pattern.

[0050] Nearest abnormal pattern recognition: Calculate the distances between the component to be tested and each abnormal pattern category in the abnormal pattern library, and identify the nearest abnormal pattern category: ; where represents the index of the abnormal pattern category that is most similar to the features of the component to be tested, represents finding the category index that makes the smallest , represents the dynamic time warping distance between the feature sequence of the component to be tested and the center of the abnormal pattern of the th category.

[0051] Abnormality degree score calculation: Calculate the normalized abnormality degree score based on the distance between the component to be tested and the nearest abnormal pattern : ; where represents the abnormality degree score of the component, with a value range of [0,1]. The closer the value is to 1, the higher the matching degree with the abnormal pattern and the greater the failure risk; is the standard deviation of the nearest abnormal pattern category, used to normalize the distance; represents the exponential function.

[0052] Confidence evaluation: Calculate the confidence of the abnormality score at the same time, comprehensively consider factors such as data integrity and abnormal pattern library coverage, and provide a reliability reference for subsequent analysis.

[0053] Through this step, an abnormality degree score is generated for each electronic component to quantify its potential failure risk, laying a foundation for subsequent failure propagation analysis.

[0054] Step 4: Construct a temporal heterogeneous graph network based on the component similarity network, system topology structure information, and component anomaly degree scores. In this step, the outputs of the previous three steps are comprehensively utilized to integrate the system topology relationship (output of Step 1), similarity relationship (output of Step 2), and anomaly degree score (output of Step 3) among electronic components into a unified graph network model to construct a temporal heterogeneous graph network. By integrating this information from different dimensions, a network model that comprehensively describes the relationships and states of component groups is formed, providing a basis for fault propagation analysis. Specifically, the system topology structure information in Step 1 is used to establish physical connection and functional dependency edges, the similarity network in Step 2 is used to construct similarity relationship edges, and the anomaly degree score in Step 3 serves as the key feature of the nodes. This step includes the following sub-steps: Step 4.1: Definition of the heterogeneous graph network structure Define the temporal heterogeneous graph network: ; Where: ; represents the set of nodes, , , respectively represent the 1st, 2nd, and th electronic components, represents the total number of electronic components; ; represents the set of edges, and the edge represents the connection relationship between the node corresponding to the component and the node corresponding to the component , where; ; represents the node feature matrix, , , respectively represent the feature vectors of the 1st, 2nd, and th electronic components, represents the total number of electronic components; ; represents the set of adjacency matrices, , and respectively represent the physical connection relationship adjacency matrix, functional dependency relationship adjacency matrix, and similarity relationship adjacency matrix; ; represents a timestamp sequence, used to record the evolution of the network structure over time; 、 、 represent the 1st, 2nd, and th observation time points respectively, represents the total number of observed time points.

[0055] Step 4.2, multi-relationship edge attribute definition; In the temporal heterogeneous graph network, three types of edge relationships are defined, representing different types of component connection relationships respectively: Physical connection relationship: Based on the system topology structure information obtained in Step 1.3, construct the physical connection relationship adjacency matrix , where represents that there is a direct physical connection between component and component , otherwise ; Functional dependency relationship: Based on the functional dependency relationship analyzed in Step 1.3, construct the functional dependency relationship adjacency matrix , where represents the degree of functional dependency of component on component ; Similarity relationship: Based on the component similarity network constructed in Step 2, form the similarity relationship adjacency matrix , where represents the similarity between component and component .

[0056] Assign weight coefficients 、 and to the edge relationships, used to balance the importance of different types of relationships, where 、 and represent the weight coefficients of the physical connection relationship, functional dependency relationship, and similarity relationship respectively.

[0057] Step 4.3, node feature vector construction; For each node in the graph network, construct its feature vector , including the following information: Basic component attributes: One-hot encoding representation of classification features such as component type, functional role, importance level, etc.; Abnormality degree score: The component abnormality degree score calculated in Step 3.3 Key features of the node; Temporal state feature: It contains the state information of components at multiple historical time points, forms a temporal state sequence, and captures the state change trend.

[0058] For different dimensions in the feature vector, normalization processing is applied to balance the dimensional differences of different features.

[0059] Step 4.4, construction of the temporal graph network model; Combining the node features and the attributes of three types of relational edges, construct a complete temporal heterogeneous graph network model: Static graph network construction: First, construct a static graph network at a single time point, and combine the node feature matrix with three types of adjacency matrices , and to form a static graph representation; Temporal extension: Expand the static graph into a temporal graph. For a series of static graph snapshots constructed for multiple historical time points, add connecting edges in the time dimension to form a temporal graph network. Specifically, for the instance of node at time points and , add directed time edges to represent the evolutionary relationship of the state over time; Attention weight calculation: Calculate attention weights for different types of relational edges and connections at different time points to highlight important node connections and key time points. For the attention weight of relational edge , it can be calculated by the following formula: ; where represents the attention weight of node to node , represents the activation function, is the weight matrix, is the attention vector, represents the transpose operator, represents the vector concatenation operation, represents the set of neighbor nodes of node , represents the exponential function, represents the summation symbol, , , respectively represent the feature vectors of nodes , , .

[0060] In practical applications, taking a certain integrated circuit board as an example, this circuit board contains various electronic components, such as processors, capacitors, resistors, etc. The temporal heterogeneous graph network model can reflect the physical connection relationships between these components (such as the data bus connection between the processor and the memory), functional dependency relationships (such as the power supply dependency of the power management chip on all devices), and similarity relationships (such as the similarity between capacitors of the same model produced in the same batch). Through this model, the interactions between components on the circuit board and potential fault propagation paths can be clearly represented and analyzed.

[0061] The output of this step is a complete temporal heterogeneous graph network model , which contains node features, multi-type edge relationships, and time dimension information, comprehensively describing the structure and state of the component system. This model will serve as the basis for predicting the fault propagation path in the next step, supporting the analysis and prediction of system-level fault risks.

[0062] Step 5: According to the temporal heterogeneous graph network, apply the graph attention network model to predict the propagation path of electronic component failures; This step is based on the temporal heterogeneous graph network model constructed in Step 4 to predict the fault propagation path in the electronic component system, identify key components and high-risk propagation links. The temporal heterogeneous graph network output in Step 4 provides the complex interaction relationships and state information between components, providing a complete data basis for the simulation and prediction of fault propagation. By applying the graph attention network and Monte Carlo tree search algorithm on this network model, the propagation mechanism and path of faults in the system can be effectively analyzed. This step includes the following sub-steps: Step 5.1: Application of the graph attention network model; Apply the graph attention network (GAT) model to analyze the influence transfer probability between nodes: Definition of the graph attention layer: Construct a multi-layer graph attention network. For the node in the th layer, its feature update formula is: ; where represents the feature representation of the node corresponding to the component in the th layer, represents the feature representation of the node corresponding to the component in the th layer, represents the set of neighbor nodes of the node corresponding to the component , represents the node corresponding to the component The attention coefficient for the component The corresponding node is, the weight matrix, is the non-linear activation function, denotes the sum of all nodes in the set of neighbor nodes of the node corresponding to the component ; Multi-head attention function: To improve the stability and expressive power of the model, the multi-head attention function is adopted. For independent attention heads, the formula for updating the node features in the -th layer is: ; where, denotes the attention coefficient of the node corresponding to the component in the -th attention head to the node corresponding to the component , denotes the weight matrix of the -th attention head, denotes the total number of attention heads, denotes the sum from 1 to for all attention heads; Heterogeneous relationship fusion: For different types of edge relationships (physical connection, functional dependency, similarity), the attention coefficients are calculated separately and then weighted fusion is performed: ; where, denotes the comprehensive influence coefficient of the component on the component , denotes the relationship type under which the component has an attention coefficient to the component , denotes the weight coefficient of the relationship type , satisfying , denotes the sum over all three relationship types, where denotes the physical connection relationship, denotes the functional dependency relationship, denotes the similarity relationship; Temporal information integration: The static graph attention model is extended to a temporal model by adding information transmission in the time dimension to capture the propagation law of faults over time.

[0063] In this embodiment, the graph attention network model specifically includes the following structure: The input layer receives the feature vectors and topological relationships of components; the hidden layer includes 3 graph attention layers, each layer uses 8 attention heads, and the activation function is ELU (Exponential Linear Unit); the output layer generates the influence transfer probability of each component on other components. This network is trained through supervised learning, using the observed fault propagation paths in historical fault cases as label data.

[0064] Taking a server motherboard as an example, when the power management chip shows abnormalities, the graph attention network model can accurately calculate the influence transfer probability of this abnormality on other components (such as the processor, memory, storage controller, etc.), so as to predict the possible fault diffusion paths. The model can identify that the abnormality of the power management chip will first affect the nearby voltage regulators and filter capacitors, then spread to core components such as the processor and memory, and may ultimately lead to the collapse of the entire system.

[0065] Step 5.2, Monte Carlo tree search simulation; Based on the calculated influence transfer probability between nodes, the Monte Carlo tree search (MCTS) method is applied to simulate the propagation process of faults in the system: Initial state definition: Select the component whose abnormality degree score exceeds the threshold as the initial fault node to form the initial state ; State transition simulation: Starting from the initial state, based on the influence transfer probability between nodes, simulate the propagation process of faults to other nodes. For the current state , the probability that the node corresponding to the component transfers to the fault state is calculated as follows: ; Where represents the probability that the node corresponding to the component transfers to the fault state, represents the set of components that are already in the fault state in state , represents the influence transfer probability of component on component , represents the product of the influence transfer probabilities on all faulty components in the set ; Monte Carlo Tree Construction: Execute multiple simulations to construct a Monte Carlo tree, recording different fault propagation paths and their probabilities. At each node of the tree, select the next action according to the UCB (Upper Confidence Bound) formula: ; where represents the average return when executing action in state , represents the average return when executing action in state , represents the number of times state is visited, represents the number of times action is executed in state , and is the exploration parameter; Path Evaluation: Evaluate the fault propagation paths obtained from the simulation, calculating indicators such as the severity, impact range, and occurrence probability of the paths.

[0066] In practical applications, the Monte Carlo tree search algorithm starts from the initial fault node (such as an abnormal capacitor), and by repeatedly simulating the process of fault propagation to other components, constructs a probability tree of fault propagation. For example, in a circuit board of a communication device, when the input filter capacitor is abnormal, the search algorithm can simulate various possible fault propagation paths, such as "filter capacitor → voltage regulator chip → signal processing chip → output driver circuit" and "filter capacitor → power distribution network → digital control unit" and other different paths, and calculate the occurrence probability and impact range of each path.

[0067] Step 5.3, Identification of Key Components and Propagation Paths; Based on the simulation results, identify the key components and high-risk propagation paths in the system: Calculation of Component Fault Sensitivity: For the node corresponding to each component i, calculate its fault sensitivity , which represents the degree of influence of this component on the overall fault state of the system: ; where represents the fault sensitivity of component i, is the total number of simulated paths, represents the th propagation path, represents the node corresponding to component i, and is an indicator function indicating whether the node ​​​It takes the value of 1 at the middle time, otherwise 0. Indicates the propagation path Degree of influence Indicates the summation from 1 to all simulated paths ; Summation Component criticality ranking: Rank components according to the calculated fault sensitivity to identify critical components in the system, which play an important role in fault propagation. High-risk propagation path identification: Based on the occurrence probability and influence range of paths, identify high-risk propagation paths in the system, which may lead to large-scale cascading failures. Visualization display: Visualize the identified critical components and high-risk propagation paths to facilitate maintenance personnel to understand the vulnerability points and potential risks of the system. Through this step, the fault propagation prediction results in the electronic component system are obtained, including information such as critical components and high-risk propagation paths, providing decision support for system maintenance and risk management.

[0068] An electronic component detection data processing system for implementing the above-mentioned electronic component detection data processing method, including: A data preprocessing module for collecting and preprocessing the detection data of multiple electronic components. A similarity analysis module for calculating the multi-dimensional similarity between multiple electronic components and establishing a component similarity network. An anomaly detection module for extracting multi-scale time series features from the component detection data, identifying anomaly patterns and calculating the component anomaly degree score. A graph network construction module for constructing a time series heterogeneous graph network. A fault propagation prediction module for applying a graph attention network model to predict the propagation path of electronic component faults.

[0069] Here, the present invention provides an implementation example: The electronic component detection data processing method of this embodiment has been practically applied on the production line of a large communication equipment manufacturing enterprise. The main control unit circuit board of the communication base station produced by this enterprise contains about 180 electronic components, including various types of components such as processors, memories, power management chips, filter capacitors, and voltage regulators. These circuit boards need to undergo strict quality inspections before leaving the factory to ensure their reliability and stability in the actual deployment environment.

[0070] Before applying this method, enterprises mainly relied on single-point independent detection methods and manual experience judgment, facing problems such as low detection efficiency, weak early fault warning ability, and difficulty in quickly discovering batch problems. By applying this method, a complete prediction system for the propagation of group faults in electronic components was constructed, realizing a comprehensive assessment of the reliability of circuit boards and early risk warning.

[0071] During the production inspection process of the main control unit circuit board of a communication base station, the following types of data were collected: Static parameters: component model, batch number, supplier code, production date, etc.; Electrical parameters: electrical characteristic parameters such as voltage, current, resistance, capacitance, etc.; Thermal parameters: operating temperature, thermal distribution, thermal resistance, etc.; Timing data: power-on startup waveform, signal integrity test data, aging test data, etc.; These data were collected by an automatic test equipment (ATE), and the sampling frequency varied according to the importance of the parameters. Taking the output voltage of a power management chip as an example, a high-frequency sampling of 10Hz was used; while for the temperature parameter, a low-frequency sampling of once per minute was used, as shown in Table 1: Table 1: Example of detection parameters for circuit board components

[0072] For the collected raw data, wavelet denoising and normalization processing were applied. Taking the output voltage of a power chip as an example, the signal-to-noise ratio after denoising was improved from the original 19.8dB to 32.4dB, significantly improving the reliability of subsequent analysis.

[0073] Based on the static attributes and dynamic performance parameters of components, the similarity between different components was calculated. Taking 10 power management chips of the same model in the same batch as an example, their feature vectors were constructed and the cosine similarity was calculated, as shown in Table 2: Table 2: Example of similarity matrix of power management chips (partial)

[0074] By setting the similarity threshold θ = 0.92 and applying the k = 3 nearest neighbor constraint, a component similarity network was constructed. In this network, similar components form clusters, providing a basis for subsequent collaborative fault prediction.

[0075] Multi-scale timing feature extraction was performed on the performance data of key components on the circuit board. Taking the leakage current timing data of a certain filter capacitor as an example, 5-layer wavelet decomposition was applied to obtain 1 group of approximation coefficients and 5 groups of detail coefficients, and the corresponding statistical features were extracted, as shown in Table 3: Table 3: Example of extraction results of timing features of filter capacitors

[0076] Based on historical fault case data, an abnormal pattern library containing 12 typical fault modes was constructed. Using the dynamic time warping algorithm, the distances between the components under test and these abnormal patterns were calculated to generate an abnormal degree score, as shown in Table 4: Table 4: Matching Degree Scores of a Certain Filter Capacitor with Abnormal Patterns

[0077] This filter capacitor has the highest matching degree with the "equivalent series resistance increasing" mode, with an abnormal score of 0.742 and a confidence level of 0.987, indicating that there are early abnormal signs of increasing ESR in this component.

[0078] Based on the design schematic diagram and measured data of the circuit board, a temporal heterogeneous graph network containing 180 nodes was constructed. The graph network contains three types of relationship edges: physical connections (such as PCB trace connections), functional dependencies (such as power supply, clock, data dependencies, etc.), and similarity relationships (such as components of the same batch and the same model).

[0079] Applying the graph attention network model, the influence transfer probability between nodes was analyzed. Taking the abnormal filter capacitor (score 0.742) as the initial node, its fault propagation path was simulated through Monte Carlo tree search, as shown in Table 5: Table 5: Simulation Results of Fault Propagation Paths (Top 5 Highest Probability Paths)

[0080] Based on the simulation results, the fault sensitivity of each component was calculated to identify key components and high-risk propagation paths, as shown in Table 6: Table 6: Ranking of Fault Sensitivity of Key Components (Top 5)

[0081] By conducting a 3-month follow-up test on 100 main control unit circuit boards of communication base stations, the prediction accuracy of this method was verified. During the test period, a total of 17 circuit boards had faults to varying degrees, which highly coincided with the high-risk circuit boards predicted by the system, as shown in Table 7: Table 7: Statistics of Fault Prediction Accuracy

[0082] This method has significantly improved in all indicators compared with traditional single-point detection methods. In particular, the early warning time has been advanced from an average of 8 hours to 72 hours, providing sufficient intervention time for maintenance personnel.

[0083] By specifically strengthening the monitoring and maintenance of the identified key components, this method significantly reduces the failure rate of the main control unit of communication base stations, as shown in Table 8: Table 8: Statistics on the effect of failure rate reduction

[0084] After applying this method, the overall failure rate of the main control unit of communication base stations has decreased from 9.10% to 5.04%, and the reduction ratio reaches 44.6%, basically achieving the expected failure rate reduction target of 45%.

[0085] In summary, the application of this embodiment in the actual communication equipment production line verifies its significant technical effects in terms of prediction accuracy, early warning time, and failure rate reduction, providing an effective technical means for ensuring the reliability of electronic components.

[0086] The above describes the embodiments of the present invention, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. A method for processing electronic component detection data, characterized in that: include: Collect and pre-process the test data of multiple electronic components to obtain high-quality component test data and system topology information; Calculate the multi-dimensional similarity between multiple electronic components based on component detection data and establish a component similarity network; Extract multi-scale time series features based on component inspection data, identify abnormal patterns and calculate component abnormality scores; Based on the component similarity network, system topology information and component anomaly degree scores, a time-series heterogeneous graph network is constructed; Based on the time-series heterogeneous graph network, the graph attention network model is applied to predict the propagation path of electronic component failures.

2. The electronic component detection data processing method according to claim 1, characterized in that: The detection data includes static parameter data, electrical parameter data, thermal parameter data and timing data.

3. The electronic component detection data processing method according to claim 1, characterized in that: The calculating of the multi-dimensional similarity between the plurality of electronic components comprises: Extract the static properties and dynamic performance characteristics of components and construct feature vectors; Cosine similarity is used to calculate the similarity between components; Set similarity thresholds and nearest neighbor constraints to build a component similarity network.

4. The electronic component detection data processing method according to claim 1, characterized in that: The multi-scale temporal feature extraction comprises: Apply wavelet transform to time series data for multi-scale decomposition; Extract statistical features from the decomposition results, including mean, variance, skewness, kurtosis, and entropy; Build an abnormal pattern library and use the dynamic time warping algorithm to calculate the distance between the components under test and the abnormal patterns; Generates component anomaly scores.

5. The electronic component detection data processing method according to claim 1, characterized in that: The temporal heterogeneous graph network comprises: Represent electronic components as nodes in a graph network; Construct three types of relationship edges, including physical connection relationships, functional dependency relationships, and similarity relationships; Integrate temporal information into the graph network to form a temporal heterogeneous graph network.

6. The electronic component detection data processing method according to claim 1, characterized in that: The graph attention network model includes: Input layer, receiving node features and edge relationship information; Multiple graph attention layers, each containing multiple attention heads, are used to learn the influence weights between nodes; Heterogeneous relationship fusion mechanism, integrating the attention coefficients of different types of edges; Timing information integration mechanism to capture the temporal dynamics of fault propagation; The output layer generates the influence transmission probability between components.

7. The electronic component detection data processing method according to claim 6, characterized in that: The heterogeneous relationship fusion mechanism is implemented by assigning different attention coefficients to different types of relationship edges and performing weighted fusion.

8. The electronic component detection data processing method according to claim 6, characterized in that: The timing information integration mechanism includes: The construction time step is The timing window of Calculate attention weights at each time step; Apply the temporal attention mechanism to integrate information from different time steps.

9. The electronic component detection data processing method according to claim 1, characterized in that: The propagation path of predicting electronic component failure includes: Application of graph attention network model: Apply graph attention network model to analyze the influence transmission probability between nodes; Monte Carlo tree search simulation, based on the calculated impact transmission probability between nodes, the Monte Carlo tree search method is used to simulate the propagation process of faults in the system; Identification of key components and propagation paths, based on simulation results, identifies key components and high-risk propagation paths in the system.

10. An electronic component detection data processing system, characterized in that: A method for processing electronic component detection data for executing any one of claims 1 to 9, comprising: A data preprocessing module, used to collect and preprocess the detection data of multiple electronic components; Similarity analysis module, used to calculate the multi-dimensional similarity between multiple electronic components and establish a component similarity network; Anomaly detection module, used to extract multi-scale time series features from component detection data, identify abnormal patterns and calculate component abnormality scores; Graph network construction module, used to build time-series heterogeneous graph networks; The fault propagation prediction module is used to apply the graph attention network model to predict the propagation path of electronic component faults.

Images