Image generation method and system applied to production of circuit board

Abstract

Embodiments of the present application disclose an image generation method and system applied to circuit board production. The method responds to a failure circuit board traceability task request, positions a failure area based on a preset failure positioning benchmark, collects a failure point morphology detail image and determines its layout position association information, and obtains a failure feature through feature mining. According to the failure feature, an abnormal association analysis is performed on a production link chain to obtain an abnormal tendency analysis result containing an association degree value of each production link and the failure feature. A potential abnormal link is locked and abnormal variable data is extracted, and an image generation rule is determined in combination with the failure feature. Finally, based on the potential abnormal link scene label and the image generation rule, an intermediate state image of the failure circuit board in the potential abnormal link is reversely generated. The method can realize accurate traceability from a failure result to a production abnormality, and present an abnormal state in a production process in a visual image, thereby realizing efficient and intuitive circuit board failure traceability.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image generation method and system for use in circuit board production. Background Technology

[0002] In the electronics manufacturing industry, circuit boards (PCBs) serve as the core components of various electronic devices, and their production quality directly determines the performance and reliability of the final products. As electronic devices become increasingly miniaturized and highly integrated, the wiring density of PCBs continues to rise, and the complexity of their processes increases. Even minor anomalies in any stage of the production process can lead to failures such as open circuits, short circuits, and voids in the via walls, resulting in significant quality cost losses for companies. Therefore, precise traceability technology for failed PCBs remains a research hotspot and an essential requirement in the PCB manufacturing field.

[0003] Currently, existing technologies in the field of PCB failure tracing mainly revolve around two directions: one is failure location technology based on electrical testing and appearance inspection, which uses online testing equipment, X-ray scanning equipment, etc., to obtain the electrical performance parameters and appearance images of the PCB, and then locates the failure area by combining it with preset failure judgment criteria; the other is anomaly tracing technology based on production data statistical analysis, which uses statistical analysis methods to identify production links that may cause failures by associating the production batch, process parameters, equipment operation data, etc. of the PCB. In addition, the application of image generation technology in the PCB manufacturing field is mostly concentrated in positive scenarios, that is, generating expected images of finished PCBs based on production process parameters for pre-production simulation and rehearsal of the production process.

[0004] However, traditional failure tracing technologies cannot accurately uncover the implicit correlation between failure characteristics and production anomalies, making it difficult to accurately trace back from the failure result to the source production process. Furthermore, existing image generation technologies cannot reverse-engineer abnormal intermediate states in the production process based on the failure result, making it difficult to provide intuitive and verifiable evidence for failure tracing. Summary of the Invention

[0005] This application provides an image generation method and system for use in circuit board production.

[0006] This application provides an image generation method for circuit board production, applied to an image generation system, the method comprising:

[0007] In response to a tracing task request for a failed circuit board, the failed circuit board is divided into regions based on a preset failure location benchmark to locate the failure area of ​​the failed circuit board.

[0008] Collect morphological detail images of the failure points within the failure area, and determine the layout location association information of the failure points in the failure circuit board through circuit board layout mapping processing;

[0009] Based on the layout location association information, feature mining processing is performed on the morphological detail image to obtain the failure features corresponding to the morphological detail image;

[0010] An anomaly correlation analysis is performed on the production chain of the failed circuit board based on the failure characteristics to obtain an anomaly tendency analysis result that includes the correlation degree value between each production link and the failure characteristics.

[0011] The abnormal tendency analysis results are used to identify potential abnormal links in the production chain and extract the abnormal variable data corresponding to the potential abnormal links. Based on the correlation and fusion processing results of the abnormal variable data and the failure features, the image generation rules are determined.

[0012] Based on the scene label of the potential abnormal link, the state of the failed circuit board in the potential abnormal link is reverse generated according to the image generation rule to obtain the intermediate state image of the failed circuit board in the potential abnormal link.

[0013] One embodiment of this application provides an image generation system, including:

[0014] A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; and when the computer program is executed by the processor, causing the processor to implement any of the image generation methods described above for use in circuit board production.

[0015] One embodiment of this application provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the image generation method applied to circuit board production.

[0016] This application's embodiments construct a complete technical chain from the global logical level, deduce production anomalies from failure results, and reconstruct intermediate states: through the collaborative processing of pre-defined failure location benchmark area division, failure point morphology detail collection, and layout location correlation information mining, it breaks through the limitation of traditional failure analysis focusing only on isolated failure phenomena, realizing the targeted identification of failure features and global correlation logic mining, and can accurately capture the intrinsic relationship between the micro-morphology of failure points and circuit board layout; based on the failure features, the production link chain anomaly correlation analysis establishes the correlation between failure features and production anomalies by quantifying the correlation degree between each production link and failure features. The implicit association mapping mechanism solves the problem of accurately locating the cause from the result in reverse tracing, effectively anchoring the source production link that caused the failure. By determining image generation rules through the association and fusion of abnormal variable data and failure features, and combining potential abnormal link scene labels to achieve reverse state generation, a scene restoration system based on production logic is constructed. This system can transform abstract production abnormal variables into visualized intermediate state images, overcoming the limitation of forward image generation that only focuses on production output. It allows abnormal states hidden in the production process to be presented intuitively, providing traceable and verifiable visualized evidence support for circuit board failure tracing. Thus, this embodiment achieves end-to-end collaboration between failure feature identification, abnormal link locking, and intermediate state image generation, improving the accuracy, efficiency, and intuitiveness of circuit board failure tracing. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an image generation method for circuit board production, provided as an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the basic structure of an image generation system provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] Please see Figure 1 , Figure 1 This is a flowchart of an image generation method for circuit board production provided in an embodiment of this application. The method can be executed by an image generation system or jointly by an image generation system and a server. The method may include steps 110-160.

[0021] Step 110: In response to the tracing task request for the failed circuit board, perform region division processing on the failed circuit board based on the preset failure location benchmark to locate the failure area of ​​the failed circuit board.

[0022] In this embodiment, after receiving a tracing task request for a failed circuit board, the image generation system retrieves a preset failure location benchmark. This benchmark includes the functional partition definition of the circuit board, the pad array layout specifications, the wiring density threshold, etc., and all benchmark parameters have undergone dimensional standardization. The image generation system acquires a global scan image of the failed circuit board through a high-precision image acquisition module. This image is stored in a lossless bitmap format and contains all the physical layout information of the circuit board. Then, the image generation system calls the region division processing module to divide the global scan image into multiple independent regions, such as the power supply area, signal transmission area, grounding area, and component soldering area, according to the functional partition definition in the failure location benchmark. Each region is marked with a unique region identifier. Further, the image generation system performs preliminary failure feature detection on each divided region, using an edge detection algorithm to identify obvious failure signs such as broken lines, detached pads, and component misalignment within the region. Simultaneously, combined with the electrical test data of the circuit board, it filters out regions with abnormal electrical pathways. Finally, areas exhibiting both explicit failure signs and abnormal electrical pathways are marked as failure areas, generating failure area location results that include the boundary coordinates of the failure areas, area identifiers, and preliminary failure types.

[0023] Step 120: Collect morphological detail images of the failure points within the failure area, and determine the layout location association information of the failure points in the failure circuit board through circuit board layout mapping processing.

[0024] In this embodiment, based on the failure area location result output in step 110, the image generation system controls a high-precision microscopic imaging unit to perform local scanning of the failure area, acquiring morphological detail images of all failure points within the failure area. The resolution of these images meets the requirement of identifying micron-level circuit defects, and the image format is a lossless bitmap containing RGB three channels. Then, the image generation system activates the circuit board layout mapping processing module, retrieving the original design layout file corresponding to the failed circuit board. This file contains structured data such as the circuit board's wiring coordinates, component package positions, and pad parameters. Further, the image generation system matches the feature points in the morphological detail images with the coordinate data in the original design layout file. A feature point matching algorithm determines the precise coordinate position of each failure point in the original design layout, while simultaneously extracting the surrounding wiring information, adjacent component types, signal transmission paths, and other related content corresponding to that position. Finally, the coordinate positions of the failure points, surrounding wiring information, adjacent component types, and signal transmission paths are integrated into layout position association information. This information is stored in a structured JSON array format, with each failure point corresponding to an independent array element.

[0025] Step 130: Perform feature mining processing on the morphological detail image based on the layout position association information to obtain the failure features corresponding to the morphological detail image.

[0026] Step 131: Based on the layout position association information, determine the effective analysis area of ​​the morphological detail image, call the multi-scale fusion analysis algorithm and input the morphological detail image into the multi-scale fusion analysis algorithm, and extract the initial features of the morphological detail image at the pixel level, texture level and structure level through feature extraction units of different scales respectively.

[0027] In this embodiment, the image generation system parses the layout position association information output in step 120, extracts data such as the surrounding wiring range and adjacent component package boundaries corresponding to the failure point, and uses this as a basis to determine the effective analysis area in the morphological detail image. The boundary of the effective analysis area is a rectangular area centered on the failure point, covering a range three times the width of the surrounding wiring. The image portion outside this area is marked as an invalid area and masked in subsequent processing. Then, the image generation system calls a preset multi-scale fusion analysis algorithm, which includes three independent modules: a pixel-level feature extraction unit, a texture-level feature extraction unit, and a structure-level feature extraction unit. The image generation system inputs a morphological detail image into the algorithm. The pixel-level feature extraction unit extracts the RGB channel values ​​of each pixel within the effective analysis area, generating a pixel-level initial feature set containing pixel location and corresponding channel values. Further, the texture-level feature extraction unit uses a gray-level co-occurrence matrix algorithm to extract texture features such as texture direction, texture contrast, and texture energy within the effective analysis area, generating a texture-level initial feature set. Finally, the structure-level feature extraction unit uses a Hough transform algorithm to extract structural features such as circuit routing, pad shape, and component pin arrangement within the effective analysis area, generating a structure-level initial feature set.

[0028] Step 132: Establish a hierarchical association index based on the layout location association information, match and associate the initial features of each scale with the hierarchical association index, and filter out candidate features associated with the layout location of the failure point through a directional identification algorithm.

[0029] In this embodiment, the image generation system establishes a hierarchical association index based on data such as the coordinates of the failure point, surrounding wiring information, and adjacent component types in the layout location association information. This index takes the failure point as the root node and has three first-level sub-nodes: wiring feature sub-node, component feature sub-node, and signal path feature sub-node. Each first-level sub-node is further subdivided into second-level sub-nodes of the corresponding type. For example, the wiring feature sub-node has sub-nodes for wiring width, wiring spacing, and wiring material. The image generation system matches and associates the pixel-level, texture-level, and structural-level initial feature sets output in step 131 with the hierarchical association index. In the pixel-level feature set, the pixel features corresponding to the wiring area are associated with the wiring feature sub-node, and the pixel features corresponding to the component area are associated with the component feature sub-node. In the texture-level feature set, the features corresponding to the wiring texture are associated with the wiring feature sub-node, and the features corresponding to the component surface texture are associated with the component feature sub-node. In the structural-level feature set, the features corresponding to the wiring direction are associated with the wiring feature sub-node, and the features corresponding to the component pin structure are associated with the component feature sub-node. The image generation system calls a directional recognition algorithm. Based on preset association rules, the algorithm filters out features that are directly related to the layout location of the failure point, such as pixel features directly connected to the wiring where the failure point is located, texture features of adjacent components corresponding to the failure point, and structural features of the signal path where the failure point is located. The features obtained above are marked as candidate features to generate a candidate feature set.

[0030] Step 133: Call the association logic mining model, input the candidate features and their corresponding layout position association information into the association logic mining model, mine the association logic topology between the candidate features through the feature association rules of the association logic mining model, and perform fusion processing on the candidate features with association relationships to obtain fused features.

[0031] In this embodiment, the image generation system calls a preset association logic mining model. This model adopts a graph neural network architecture and includes four core modules: an input layer, a feature association layer, a topology generation layer, and a feature fusion layer. The image generation system inputs the candidate feature set output in step 132 and the corresponding layout position association information into the input layer of the model. The input layer standardizes the candidate features to unify the dimensions of different types of features, and converts the layout position association information into association weight parameters that the model can recognize. The feature association layer mines the inherent association relationships between candidate features based on preset feature association rules. For example, the association relationship between line break edge pixels in pixel-level features and wiring direction features in structural-level features, and the association relationship between component surface oxidation texture in texture-level features and component type in layout position association information. Then, the topology generation layer generates an association logic topology based on the mined association relationships. This topology uses each candidate feature as a node and the association relationships between features as edges, with the edge weights corresponding to the tightness of the association relationships. The feature fusion layer fuses candidate features that are related. It merges multiple related features into a high-dimensional feature vector by concatenating them along the feature dimension. Each high-dimensional feature vector corresponds to a set of related features, generating a fused feature set.

[0032] Step 134: Based on the aforementioned association logic topology, generate a feature association confidence assessment strategy, quantify the credibility of the association logic corresponding to the fused feature, and generate feature association confidence parameters.

[0033] In this embodiment, the image generation system generates a feature association confidence evaluation strategy based on the association logic topology output in step 133. This strategy includes three core rules: association path length evaluation rule, association node number evaluation rule, and association edge weight sum evaluation rule. Specifically, the association path length evaluation rule shows a negative correlation between association path length and confidence, meaning a shorter path indicates higher confidence; the association node number evaluation rule shows a positive correlation between the number of association nodes and confidence, meaning a larger number of nodes indicates higher confidence; and the association edge weight sum evaluation rule shows a positive correlation between the sum of association edge weights and confidence, meaning a higher sum of weights indicates higher confidence. The image generation system performs quantitative evaluation according to the three evaluation rules for the association logic topology corresponding to each fusion feature. The evaluation results of each rule are then weighted and fused. The weighting coefficients are determined based on weight parameters trained in advance using historical data. Finally, a feature association confidence parameter is generated for each fusion feature. This parameter is a value between 0 and 1; a higher value indicates higher confidence in the association logic of the corresponding fusion feature. All confidence parameters of fusion features are bound and stored with the fusion feature set.

[0034] Step 135: Create a dynamic weight allocation model by combining the layout location association information, take the feature association confidence parameter as input, and assign dynamic weights to the information of different association dimensions in the fused features through the dynamic weight allocation model.

[0035] In this embodiment, the image generation system combines data such as the failure point location, surrounding wiring type, and importance of adjacent components from the layout location association information to create a dynamic weight allocation model. This model adopts a multilayer perceptron neural network architecture. The input layer consists of feature association confidence parameters and key parameters from the layout location association information, while the output layer consists of weight values ​​for different association dimensions in the fused features. The image generation system inputs the feature association confidence parameters and key parameters from the layout location association information output in step 134 into the input layer of the model. The input layer standardizes the data and then passes it to the hidden layer. The hidden layer performs a non-linear transformation on the data using a preset activation function to uncover the intrinsic mapping relationship between the feature association confidence parameters and the layout location association information. Then, based on the output of the hidden layer, the output layer assigns dynamic weights to the pixel dimension features, texture dimension features, and structural dimension features in the fused features. The weight allocation rule is: the higher the feature association confidence parameter of the fused feature, the higher the weight of its corresponding core association dimension; the greater the impact of the failure in the layout location association information, the higher the weight of the corresponding feature. For example, if the failure point is located in the signal transmission area and the feature association confidence parameter is high, then the signal path feature weight in the structural dimension feature will be assigned a higher value. Ultimately, each different association dimension of the fused feature corresponds to a dynamic weight value, the sum of which is 1, and all dynamic weight values ​​will be bound and stored with the fused feature set.

[0036] Step 136: Perform hierarchical enhancement processing on the fusion feature based on the dynamic weight and establish a dynamic mapping relationship between the fusion feature and the layout position association information.

[0037] In this embodiment, the image generation system performs hierarchical enhancement processing on different related dimension features in the fusion feature according to the dynamic weight value output in step 135 for each fusion feature. Specifically, the feature vector of each dimension is multiplied element by element with the corresponding dynamic weight value to obtain the enhanced feature vectors of each dimension. Then, the enhanced feature vectors of each dimension are concatenated along the feature dimensions to obtain the fusion feature after hierarchical enhancement.

[0038] The image generation system establishes a dynamic mapping relationship between the layered enhanced fusion features and the layout location association information. Specifically, it adds metadata tags such as the coordinates of the failure point, the region identifier, and the type of adjacent components from the layout location association information to each fusion feature. At the same time, it constructs a mapping index table, which contains the unique identifier of the fusion feature, the storage address of the corresponding layout location association information, the dynamic weight value, and other contents. The layout location association information corresponding to any fusion feature can be quickly queried through this index table.

[0039] Step 137: Based on the dynamic mapping relationship, the fusion features after layered enhancement are associated and integrated to obtain the failure features corresponding to the morphological detail image. The failure features contain dual attributes of morphological features and layout association features, and each failure feature carries a corresponding layout position association marker and feature association confidence parameter.

[0040] In this embodiment, the image generation system, based on the dynamic mapping relationship established in step 136, integrates all the layered enhanced fusion features. Multiple fusion features corresponding to the same failure point are aggregated through a mapping index table. The aggregation method involves concatenating the vectors of multiple fusion features along the feature dimension to obtain a high-dimensional feature vector containing all associated features of the failure point. Further, a corresponding layout location association marker is added to each aggregated feature vector. This marker includes the failure point coordinates, region identifier, signal path name, etc., and is bound to the feature association confidence parameter generated in step 134. Finally, all aggregated feature vectors are integrated into a failure feature set. Each failure feature contains dual attributes: morphological features (pixels, texture, structural dimension) and layout association features (position, surrounding wiring, adjacent components). Each failure feature carries a unique layout location association marker and a corresponding feature association confidence parameter.

[0041] Step 140: Perform anomaly correlation analysis on the production chain of the failed circuit board based on the failure characteristics to obtain anomaly tendency analysis results that include the correlation degree values ​​between each production link and the failure characteristics.

[0042] Step 141: Perform structured encoding on the failure features to obtain a structured failure feature dataset. Decompose the attributes of each production link in the production chain, extract the process parameters, operating procedures and material information of each production link, and generate a production link attribute dataset.

[0043] In this embodiment, the image generation system performs structured encoding on the failure feature set output in step 137. Specifically, it converts the morphological features, layout association features, layout position association markers, and feature association confidence parameters of each failure feature into a format that can be stored in a structured database. The morphological features are stored as floating-point arrays, the layout association features are stored as a combination of strings and numbers, the layout position association markers are stored as structured JSON objects, and the feature association confidence parameters are stored as floating-point numbers. All encoded failure features are integrated into a structured failure feature dataset and stored in the corresponding table of a relational database. Then, the image generation system retrieves the production chain information corresponding to the failed circuit board. This information includes all production stages from raw material preparation, substrate manufacturing, circuit printing, component soldering, performance testing to finished product packaging. The image generation system decomposes the attributes of each production stage, extracting the core process parameters, operating specifications, and key information of the materials used in each stage. The process parameters are stored as numerical arrays, the operating specifications are stored as a set of strings, and the key material information is stored as a structured JSON object. All extracted content is integrated into a production stage attribute dataset and stored in the same database as the structured failure feature dataset.

[0044] Step 142: Generate a latent association graph based on the structured failure feature dataset and the production process attribute dataset. Use the failure features and each production process as nodes of the latent association graph, and use the potential association between the failure features and each production process as edges of the latent association graph.

[0045] In this embodiment, the image generation system retrieves a preset association rule base, which contains definitions of the associations between each production stage and common failure features in the circuit board manufacturing process. Examples include the association between substrate manufacturing and short-circuit failure, and the association between component soldering and pad detachment failure. Then, based on the association rule base, the image generation system matches each failure feature in the structured failure feature dataset with each production stage in the production stage attribute dataset to identify potential associations. For example, if the failure feature is a short circuit, it matches the substrate manufacturing stage, circuit printing stage, and other production stages that may cause this failure. Further, the image generation system generates a latent association graph, stored in an adjacency list format. Each failure feature and each production stage is treated as an independent node, with each node containing node type (failure feature or production stage), node identifier, and node attributes. The identified potential associations are treated as edges, with each edge containing a starting node identifier, an ending node identifier, and the source of the association rule.

[0046] Step 143: Input the implicit association graph into a preset graph neural network model, perform feature embedding processing on each node through the node embedding layer of the graph neural network model, and use a multi-hop inference algorithm to mine the multi-order association relationship between the failure feature node and the production link node in the implicit association graph.

[0047] Step 1431: Perform preprocessing on the latent association graph, including outlier node removal and feature dimension supplementation, to obtain a preprocessed latent association graph.

[0048] In this embodiment, the image generation system removes outliers from the latent association graph. It counts the number of connecting edges for each node using a node degree calculation method, marking nodes with zero connecting edges as outliers and removing them from the graph. These outliers are typically production processes or isolated failure features that have no potential connection to the current failure feature. Furthermore, the image generation system supplements the feature dimensions of the remaining nodes. For failure feature nodes, it supplements their corresponding feature association confidence parameters, layout location association information, etc.; for production process nodes, it supplements their corresponding production time sequence information, upper and lower limits of process parameters, historical anomaly records, etc., ensuring that the feature dimensions of all nodes are consistent and the information is complete. The graph after outlier removal and feature dimension supplementation is stored in the graph database as a preprocessed latent association graph.

[0049] Step 1432: Input the preprocessed latent association graph into a preset graph neural network model, call the node embedding layer of the graph neural network model, use an attention-based node embedding algorithm to encode the features of each node, fuse the inherent attribute features of each node with the graph topology features, and generate an initial node embedding vector.

[0050] In this embodiment, the image generation system inputs the preprocessed latent association graph into a preset graph neural network model. This model adopts a graph attention network architecture, comprising four core modules: a node embedding layer, an attention calculation layer, a feature fusion layer, and an output layer. The image generation system calls the node embedding layer, employing an attention-based node embedding algorithm to encode the features of each node. The algorithm converts the inherent attribute features of each node into a high-dimensional vector, and simultaneously generates neighbor association vectors for each node based on the topological features of the graph. Then, it calculates the attention weights of each node and its neighboring nodes using an attention mechanism. The calculation of attention weights is based on the similarity of the node's inherent attribute features and the tightness of the topological connections. The inherent attribute feature vectors of the nodes are fused with the neighbor association vectors weighted by the attention weights. The fusion method involves concatenating along the feature dimension to generate an initial node embedding vector for each node. This vector contains both the inherent attribute information of the node and the topological information of the graph. All initial node embedding vectors are stored in the model's intermediate data cache.

[0051] Step 1433: Construct a node feature similarity matrix based on the initial node embedding vector, solve the feature matching degree between each node by using the cosine similarity calculation method, use the feature matching degree as a constraint condition for node embedding optimization, and iteratively optimize the initial node embedding vector until the difference value of the node embedding vector generated by two adjacent iterations is less than a preset difference threshold, and obtain the optimized node embedding vector.

[0052] In this embodiment, the image generation system constructs a node feature similarity matrix based on the initial node embedding vector output in step 1432. Each element in the matrix corresponds to the cosine similarity value between two nodes. The cosine similarity is calculated by dividing the inner product of the node embedding vectors by the product of their magnitudes. The value is between -1 and 1, with a higher value indicating a higher feature matching degree between the two nodes. The image generation system uses the node feature similarity matrix as a constraint to iteratively optimize the initial node embedding vector. The optimization algorithm uses gradient descent. In each iteration, the value of the node embedding vector is adjusted according to the node feature similarity matrix, making the node embedding vectors with higher feature matching degrees more similar and the node embedding vectors with lower feature matching degrees more different. After each iteration, the difference between the currently generated node embedding vector and the vector generated in the previous iteration is calculated. The difference is calculated using the average Euclidean distance of all node vectors. When the difference is less than a preset difference threshold, the iteration stops, and the current node embedding vector is used as the optimized node embedding vector. All optimized node embedding vectors replace the initial node embedding vectors and are stored in the intermediate data cache.

[0053] Step 1434: Start the preset multi-hop inference algorithm, set the inference hop count threshold based on the optimized node embedding vector, take each failed feature node as an independent starting node, and perform a hop-by-hop traversal according to the edge connection relationship of the preprocessed implicit association graph to mine the first-order association production link node directly connected to the starting node. Then, take the mined first-order association production link node as the intermediate transmission node and continue to traverse and mine the second-order association node connected to the intermediate transmission node, and so on until the preset inference hop count threshold is reached.

[0054] In this embodiment, the image generation system initiates a preset multi-hop inference algorithm. Based on the optimized node embedding vectors, it calculates the correlation strength of each node. The correlation strength is determined by the mean cosine similarity of the node embedding vectors. A threshold for the number of inference hops is set according to the correlation strength; the higher the correlation strength, the larger the threshold, thus enabling the discovery of all potential related nodes. The image generation system uses each failure feature node as an independent starting node and performs a hop-by-hop traversal according to the edge connections of the preprocessed implicit correlation graph. It discovers production link nodes directly connected to the starting node and marks them as first-order related production link nodes. Using each first-order related production link node as an intermediate transit node, it traverses and discovers other nodes connected to that node (including other production link nodes and failure feature nodes), marking them as second-order related nodes. Using each second-order related node as an intermediate transit node, it continues to traverse and discover third-order related nodes, and so on, until the number of hops reached the preset inference hop threshold. During the traversal process, the traversal result of each node is recorded, including the identifier of the related node, the number of hops in the related path, and the corresponding edge information.

[0055] Step 1435: During the multi-hop inference process, combining the correlation of process parameters in the production link attribute dataset and the feature correlation attributes in the structured failure feature dataset, dynamic correlation weights are assigned to the correlation relationships obtained by each hop mining. The dynamic correlation weights are positively correlated with the feature matching degree between nodes and the degree of process correlation.

[0056] In this embodiment, at each stage of multi-hop inference, the image generation system combines the correlation of process parameters in the production process attribute dataset and the feature correlation attributes in the structured failure feature dataset to assign dynamic correlation weights to each obtained correlation. For the two nodes corresponding to each correlation, the correlation of process parameters between the two nodes is extracted from the production process attribute dataset. For example, if the two nodes are the substrate manufacturing process and the circuit printing process, their correlation is the degree of correlation between substrate flatness and circuit printing accuracy. The correlation of feature attributes between the two nodes is extracted from the structured failure feature dataset. For example, if one node is a failure feature node and the other is a production process node, their correlation attribute is the historical probability that the production process leads to the corresponding failure feature. The correlation of process parameters, the correlation attributes, and the node feature matching degree obtained in step 1433 are weighted and fused. The weighting coefficient is determined according to the weight parameters trained in advance based on historical data. Finally, a dynamic correlation weight is generated for each correlation. This weight is a value between 0 and 1. The higher the value, the stronger the correlation. All dynamic correlation weights are bound to the corresponding correlation and stored.

[0057] Step 1436: After hierarchically marking the association relationships corresponding to different hop counts, generate a multi-level association relationship graph by combining the association nodes, association edges and corresponding dynamic association weights of each level.

[0058] In this embodiment, the image generation system hierarchically labels all relationships obtained during multi-hop inference. The labeling includes information such as the hop count (first-order, second-order, third-order, etc.), start node identifier, and end node identifier. The image generation system combines the associated nodes, edges, and corresponding dynamic association weights at each level to generate a multi-order association graph. This graph is stored in an adjacency list format. Each node includes a node identifier, node type, and node embedding vector; each edge includes a start node identifier, end node identifier, hierarchical label, and dynamic association weight. Unlike the preprocessed implicit association graph, the multi-order association graph includes indirect relationships obtained through multi-hop inference, such as the association between failure feature nodes and third-order associated production process nodes, thus more comprehensively reflecting the intrinsic connection between failure features and production processes.

[0059] Step 1437: Extract the association path information between each failure feature node and the production link node based on the multi-level association graph. Quantify and statistically analyze the length of each association path and the sum of the dynamic association weights of each association edge on the path according to the association path information, and filter out the valid association paths whose sum of dynamic association weights exceeds the preset weight threshold.

[0060] In this embodiment, the image generation system is based on a multi-level association graph. Starting with each failure feature node and ending with each production process node, it extracts all association path information between them. Each association path includes the identifiers of all nodes and all associated edges, the path length (number of hops), and the dynamic association weights of each associated edge. Then, the image generation system quantifies and statistically analyzes the length of each association path and the sum of the dynamic association weights of each associated edge. The sum of the dynamic association weights is calculated by adding the dynamic association weights of all associated edges on the path. Further, the image generation system compares the sum of the dynamic association weights with a preset weight threshold, which is trained based on historical anomaly data. Association paths whose sum of dynamic association weights exceeds the preset weight threshold are marked as valid association paths. These paths indicate a relatively close intrinsic relationship between the failure feature and the production process, and all valid association paths are stored separately.

[0061] Step 1438: Combine the effective association paths obtained from the screening with the corresponding node association information and weight information to generate a multi-level association relationship between the failure feature nodes and the production process nodes, including association level, association strength and association path.

[0062] In this embodiment, the image generation system combines the filtered valid association paths with corresponding node association information and weight information to generate multi-level association relationships between failure feature nodes and production process nodes. For each valid association path, its association level (the number of hops in the corresponding path), association strength (the sum of dynamic association weights in the corresponding path), and complete association path information (including information on all nodes and edges) are extracted. Then, all valid association paths between the same failure feature node and the same production process node are aggregated, retaining the path with the highest association strength as the core association path, while recording the association level and association strength of all association paths. Finally, the association level, association strength, and association path information between all failure feature nodes and production process nodes are integrated into a multi-level association relationship dataset. Each entry includes a failure feature node identifier, a production process node identifier, an association level, an association strength, a core association path, and a list of all association paths.

[0063] Step 144: Based on the multi-level correlation relationship, analyze the causal correlation strength between each production link and the failure feature through the causal inference algorithm, identify the collaborative anomaly patterns corresponding to different combinations of production links and failure features, and quantify the correlation strength between each production link and the failure feature to obtain the correlation degree value.

[0064] In this embodiment, the image generation system analyzes the causal correlation strength between each production link and the failure feature based on the multi-level correlation dataset output in step 1438 using a causal inference algorithm. For each combination of production link node and failure feature node, the algorithm identifies possible causal transmission paths based on the correlation path information in the multi-level correlation and excludes correlations that only have correlation but no causation. Then, using the propensity score matching method, the abnormal data of the production link is matched with the occurrence probability of the failure feature, and the causal correlation strength between the two is calculated. This strength is a value in the range of 0 to 1, and the higher the value, the stronger the causal correlation.

[0065] The image generation system identifies collaborative anomaly patterns corresponding to different combinations of production processes and failure features. Through association rule mining algorithms, it mines the probability increase of failure features when multiple production processes simultaneously exhibit anomalies. Combinations of production processes with probability increases exceeding a preset threshold are marked as collaborative anomaly patterns. Each collaborative anomaly pattern includes a production process combination identifier, a failure feature identifier, and a probability increase. Finally, the image generation system quantifies the correlation strength between each production process and the failure feature to obtain a correlation degree value. This value is a weighted fusion result of the causal correlation strength and the contribution value of the collaborative anomaly pattern. The contribution value of the collaborative anomaly pattern is the average probability increase of all collaborative anomaly patterns involving that production process. The weighting coefficients are determined based on weight parameters pre-trained using historical data. All correlation degree values ​​are bound and stored with their corresponding production processes and failure feature combinations.

[0066] Step 145: Integrate the correlation degree value, collaborative anomaly pattern and corresponding production process information to obtain an anomaly tendency analysis result that includes the correlation degree value of each production process and failure characteristics.

[0067] In this embodiment, the image generation system integrates the correlation degree value, collaborative anomaly mode, and corresponding production process information output in step 144. It creates an analysis result entry for each combination of production process and failure feature, including the production process identifier, failure feature identifier, correlation degree value, causal correlation strength, and participation status of the collaborative anomaly mode. Furthermore, it creates an independent entry for each collaborative anomaly mode, including the collaborative anomaly mode identifier, a list of participating production processes, a list of corresponding failure features, and the probability increase. Finally, it integrates all entries into an anomaly tendency analysis result, which is stored in structured JSON format. Simultaneously, it generates a corresponding visualization report, including a correlation degree value heatmap and a collaborative anomaly mode topology map.

[0068] Step 150: Based on the anomaly tendency analysis results, identify potential abnormal links in the production chain and extract the abnormal variable data corresponding to the potential abnormal links. Determine the image generation rules based on the correlation and fusion processing results of the abnormal variable data and the failure features.

[0069] Step 151: Extract the correlation degree value from the abnormal tendency analysis results, and identify the production links with a correlation degree exceeding the preset correlation degree threshold as potential abnormal links in the production link chain.

[0070] In this embodiment, the image generation system extracts the correlation values ​​between all production links and failure feature combinations from the anomaly tendency analysis results. The preset correlation threshold is trained based on historical anomaly data and is a value between 0 and 1. The higher the value, the stronger the correlation between the selected potential anomaly links and failure features. Then, the image generation system compares the average correlation value of each production link (i.e., the average correlation value between the production link and all failure feature combinations) with the preset correlation threshold. Production links with an average correlation value exceeding the threshold are marked as potential anomaly links in the production link chain. If a production link participates in a collaborative anomaly mode, and the probability increase of this mode exceeds a preset magnitude threshold, it will also be marked as a potential anomaly link even if its average correlation value does not exceed the threshold. This locks down all anomaly links that may cause failures. All marked potential anomaly links are integrated into a potential anomaly link list, which includes the production link identifier, correlation value, participation in collaborative anomaly modes, etc.

[0071] Step 152: Extract abnormal variable data from the potential abnormal links based on the preset variable extraction template. The abnormal variable data includes process fluctuation data, equipment operating parameter deviation data, environmental variable change data, and material characteristic deviation data. Standardize the extracted abnormal variable data to obtain standardized abnormal variable data.

[0072] In this embodiment, the image generation system extracts abnormal variable data from potential abnormal processes based on a preset variable extraction template. The template pre-defines corresponding variable extraction rules for different types of production processes. For example, for the component soldering process, extracted variables include soldering temperature fluctuation data, pick-and-place machine operating parameter deviation data, workshop temperature and humidity change data, and solder wire characteristic deviation data. The image generation system retrieves production process data from potential abnormal processes. This data is stored in the production process database and includes real-time collected process parameters, equipment operating parameters, environmental parameters, and material detection data. Following the rules of the variable extraction template, the system extracts corresponding process fluctuation data (such as deviations between process parameters and set values), equipment operating parameter deviation data (such as deviations between equipment speed and current and rated values), environmental variable change data (such as deviations between temperature and humidity and standard values), and material characteristic deviation data (such as deviations between material thickness and resistivity and nominal values). The image generation system standardizes the extracted outlier data by using the Z-Score standardization method to convert all variable data into standardized data with a mean of 0 and a standard deviation of 1. This eliminates the dimensional differences between different variables, ensures the accuracy of subsequent correlation and fusion processing, and finally generates standardized outlier data, which is stored in a floating-point array format.

[0073] Step 153: Create an association fusion processing model. Input the standardized abnormal variable data and the failure features into the association fusion processing model. Use the dynamic association analysis unit in the association fusion processing model to mine the dynamic association relationship between the standardized abnormal variable data and the failure features. Determine the causal association path between the abnormal variable data and the failure features based on the causal reasoning algorithm.

[0074] In this embodiment, the image generation system creates an association fusion processing model. This model adopts a deep neural network architecture and includes three core modules: a dynamic association analysis unit, a causal inference unit, and a fusion output unit. The image generation system inputs the standardized abnormal variable data output in step 152 and the failure features output in step 137 into the model's dynamic association analysis unit. This unit uses an attention mechanism to mine the dynamic association between the standardized abnormal variable data and the failure features. By calculating the attention weight for each abnormal variable and each failure feature, the degree of association between them is determined. The attention weight is calculated based on the similarity of the feature vectors of the variable and the feature, and the weight parameters obtained from training with historical association data. Then, the dynamic association analysis unit inputs the obtained dynamic associations into the causal inference unit. This unit determines the causal association path between the abnormal variable data and the failure features based on a causal inference algorithm. The algorithm excludes associations that only have correlation but no causation. The causality of each association is verified through counterfactual reasoning. If a change in the abnormal variable leads to a significant change in the probability of the failure feature occurring, then a causal association is considered to exist between the two. Further, the transmission path of the causal association is determined through path analysis. For example, deviations in equipment operating parameters lead to process fluctuations, which in turn lead to circuit printing defects, ultimately triggering failure features. All identified causal paths will be recorded, including the initial abnormal variable, intermediate transmission links, termination failure characteristics, and causal relationship strength.

[0075] Step 154: Input the dynamic correlation and the causal correlation path into the logic transformation module. The logic transformation module transforms the dynamic correlation and the causal correlation path into spatial transformation parameters, pixel distribution rules and color mapping standards required for image generation. Combine the preset visualization rules to optimize the spatial transformation parameters, pixel distribution rules and color mapping standards to generate the image generation rules. The image generation rules include the resolution of the generated image, the pixel arrangement and the feature visualization mapping relationship.

[0076] In this embodiment, the image generation system inputs the dynamic correlation and causal correlation path output in step 153 into the logic transformation module. This module comprises three core parts: a parameter transformation unit, a rule optimization unit, and a rule output unit. The parameter transformation unit converts the dynamic correlation into pixel distribution rules. The rules include the pixel distribution position and density of abnormal variables corresponding to different failure characteristics in the image; for example, abnormal variables with higher correlation correspond to higher pixel density. The causal correlation path is converted into spatial transformation parameters. The parameters include the direction, amplitude, and region of process fluctuation propagation in the image; for example, a longer causal correlation path corresponds to a larger amplitude of spatial transformation. Simultaneously, the parameter transformation unit generates a color mapping standard based on the strength of the dynamic correlation. Higher correlation strength corresponds to higher color saturation, and different types of abnormal variables correspond to different base colors; for example, process fluctuation data corresponds to blue, and equipment operating parameter deviation data corresponds to red.

[0077] The rule optimization unit optimizes spatial transformation parameters, pixel distribution rules, and color mapping standards by combining preset visualization rules. These visualization rules include requirements for image clarity, color harmony, and feature recognition. For example, optimizing pixel distribution rules ensures that pixel regions of different abnormal variables do not overlap excessively, and optimizing color mapping standards ensures that color differences are sufficiently obvious and conform to human visual habits. Finally, the rule output unit integrates the optimized spatial transformation parameters, pixel distribution rules, and color mapping standards into image generation rules. It also adds information such as the resolution of the generated image, pixel arrangement, and feature visualization mapping relationships, enabling the generated image to accurately represent the correlation between abnormal variables and failure features in potential abnormal processes. The image generation rules are stored in structured JSON format.

[0078] Step 160: Based on the scene label of the potential abnormal link, perform reverse generation processing on the state of the failed circuit board in the potential abnormal link according to the image generation rules to obtain the intermediate state image of the failed circuit board in the potential abnormal link.

[0079] Step 161: Based on the scene label of the potential abnormal link, call the corresponding scene restoration model, convert the image generation rule into the input parameters of the scene restoration model, and simultaneously input the abnormal variable data of the potential abnormal link and the failure characteristics into the scene restoration model.

[0080] In this embodiment, the image generation system extracts scene labels for each potential abnormal step from the list of potential abnormal steps. Scene labels include production step type, process type, core equipment type, etc. For example, the scene label for the component welding step is "welding-reflow soldering-pick-up machine". Then, the image generation system calls the corresponding scene reconstruction model based on the scene label. Each scene label corresponds to a pre-trained scene reconstruction model, which uses a generative adversarial network architecture and is optimized for different types of production steps. Further, the image generation system transforms the image generation rules output in step 154 ​​into input parameters for the scene reconstruction model. These parameters include spatial transformation parameters, pixel distribution rules, color mapping standards, resolution requirements, etc., and the parameter format is consistent with the input layer requirements of the model. Simultaneously, the image generation system inputs the abnormal variable data and failure features of the potential abnormal steps into the scene reconstruction model. The abnormal variable data uses the standardized abnormal variable data output in step 152, and the failure features use the failure feature set output in step 137, enabling the model to obtain the complete information needed to generate intermediate state images.

[0081] Step 162: Activate the inverse generation model through the generative adversarial network in the scene restoration model, and start the physical process simulation unit based on the current state image of the failed circuit board and the input parameters to simulate the physical change process of the failed circuit board in the potential abnormal link and generate an initial intermediate state image.

[0082] Step 1621: Invoke the preset generative adversarial network in the scene restoration model, initialize the generator and discriminator of the generative adversarial network, use the failure features and the abnormal variable data of the potential abnormal links as the initial input data of the generator, use the pre-collected state image dataset of the failed circuit board in the normal production process as the training sample set of the discriminator, and iteratively train the generative adversarial network through the gradient descent algorithm until the image generated by the generator passes the authenticity verification of the discriminator and the verification pass rate reaches the preset pass rate threshold.

[0083] In this embodiment, the image generation system calls the pre-defined generative adversarial network (GAN) in the scene reconstruction model to initialize the generator and discriminator. The generator adopts a convolutional neural network architecture, including an input layer, multi-scale convolutional layers, a feature fusion layer, and an output layer. The discriminator also adopts a convolutional neural network architecture, including an input layer, a feature extraction layer, a classification layer, and an output layer. The image generation system uses failure features and abnormal variable data of potential abnormal links as the initial input data of the generator, with the data format being high-dimensional feature vectors. A pre-collected dataset of images of the state of failed circuit boards during normal production is used as the training sample set for the discriminator. This dataset contains normal state images of different production stages, and all images have undergone preprocessing such as grayscale correction and noise filtering. Then, the image generation system iteratively trains the GAN using the gradient descent algorithm. In each iteration, the generator generates simulated abnormal state images based on the input data, and the discriminator judges the realism of the generated images and feeds back the judgment results to the generator. The generator adjusts the network parameters based on the feedback results to improve the realism of the generated images. When the image generated by the generator passes the discriminator's authenticity verification and the verification pass rate reaches the preset pass rate threshold, training stops and the current generator and discriminator parameters are saved.

[0084] Step 1622: Activate the corresponding inverse generation model based on the trained generative adversarial network, input the current state image of the failed circuit board into the inverse generation model, and perform grayscale correction, noise filtering and edge enhancement on the current state image through the image preprocessing unit of the inverse generation model to obtain a standardized current state image.

[0085] In this embodiment, the image generation system activates the corresponding inverse generation model based on the trained generative adversarial network. This model comprises four core parts: an image preprocessing unit, a parameter matching module, a physical process simulation unit, and an image output unit. The image generation system inputs the current state image of the failed circuit board into the image preprocessing unit of the inverse generation model. The current state image is a magnified view of the failure area of ​​the global scan image obtained in step 110. The preprocessing unit performs grayscale correction on the current state image, using histogram equalization to adjust the grayscale distribution and ensure uniform brightness. Further noise filtering is performed, using a Gaussian filter algorithm to remove random noise while preserving edge features. Finally, edge enhancement is performed, using the Laplacian operator algorithm to enhance edge information and make the failure features clearer. After processing, a standardized current state image is obtained, whose resolution, grayscale distribution, and edge sharpness all meet the requirements for subsequent processing.

[0086] Step 1623: Decompose the input parameters transformed by the image generation rules into spatial configuration parameters, physical simulation parameters, and feature mapping parameters. Simultaneously input the standardized current state image and the decomposed input parameters into the parameter matching module of the inverse generation model to establish a one-to-one mapping relationship between the feature points of the standardized current state image and each input parameter.

[0087] In this embodiment, the image generation system decomposes the input parameters transformed from image generation rules into spatial configuration parameters, physical simulation parameters, and feature mapping parameters. Spatial configuration parameters include image resolution, pixel arrangement, and region division rules; physical simulation parameters include spatial transformation amplitude, conduction direction, and simulation step size; and feature mapping parameters include color mapping standards, feature pixel density, and correlation strength of abnormal variables. Then, the image generation system synchronously inputs the standardized current-state image and the decomposed input parameters into the parameter matching module of the inverse generation model. The module identifies key feature points in the standardized current-state image using a feature point detection algorithm. These feature points include failure point locations, pad locations, and line intersections. Furthermore, a one-to-one mapping relationship is established between each feature point and its corresponding input parameter. For example, the failure point location corresponds to the center coordinates of the abnormal region in the spatial configuration parameters, the pad location corresponds to the conduction start point in the physical simulation parameters, and the line intersection corresponds to the high correlation strength region in the feature mapping parameters, ensuring that the input parameters accurately apply to the corresponding positions in the image.

[0088] Step 1624: Based on the one-to-one mapping relationship, start the physical process simulation unit, and construct a targeted physical field coupling simulation model in combination with the process type of the potential abnormal link. The physical field coupling simulation model is used to simulate the synergistic effect of temperature field, stress field and material action field on the failed circuit board.

[0089] In this embodiment, the image generation system initiates the physical process simulation unit based on the one-to-one mapping relationship established in step 1623. The unit constructs a targeted physical field coupling simulation model based on the process type of potential abnormal links. For example, for the component soldering process, a coupled simulation model of temperature field, stress field, and solder flow field is constructed; for the circuit printing process, a coupled simulation model of ink flow field, electric field, and drying temperature field is constructed. The physical field coupling simulation model contains multiple independent physical field sub-models, each simulating the effect of a physical field. Data interaction between sub-models is achieved through preset coupling rules. For example, changes in the temperature field lead to changes in the stress field distribution, and changes in the stress field lead to changes in the material distribution. The boundary conditions of the simulation model are determined based on the abnormal variable data of potential abnormal links. For example, equipment operating parameter deviation data corresponds to temperature deviation in the temperature field, and environmental variable change data corresponds to external stress changes in the stress field, ensuring that the simulation results reflect the actual effect of abnormal variables on the circuit board.

[0090] Step 1625: Transform the process fluctuation data and equipment operating parameter deviation data in the abnormal variable data into the boundary constraints of the physical field coupling simulation model, simulate the dynamic physical environment change process of the failed circuit board in the potential abnormal link, and at the same time, combine the morphological details in the failure characteristics to simulate the deformation, oxidation and circuit connection relationship changes of the circuit board material under the action of the dynamic physical environment.

[0091] In this embodiment, the image generation system transforms process fluctuation data and equipment operating parameter deviation data from abnormal variable data into boundary constraints of the physical field coupling simulation model. For example, it transforms welding temperature fluctuation data into time series variation constraints of the temperature field and pick-and-place machine operating parameter deviation data into initial stress constraints of the stress field. The physical field coupling simulation model simulates the dynamic physical environment changes of the failed circuit board in the potential abnormal stage, performing iterative calculations according to a preset simulation step size. Each step calculates the distribution data of the physical field, including temperature distribution, stress distribution, and material distribution. Simultaneously, the model combines the morphological details in the failure features to simulate the deformation, oxidation, and changes in circuit connection relationships of the circuit board material under the action of the dynamic physical environment. For example, if the failure feature is circuit breakage, it simulates the deformation process of the circuit under the action of the stress field until the breakage state is reached; if the failure feature is pad oxidation, it simulates the oxidation process of the pad material under the action of the temperature field and the ambient humidity field until an oxidation morphology matching the failure feature is generated.

[0092] Step 1626: In the physical change process simulation stage, the simulation process is discretized according to the preset time step, and the morphological parameters, material performance parameters and surrounding environment parameters of the failed circuit board are recorded in real time at each time step. The parameter data corresponding to each time step is converted into pixel raw data that meets the requirements of image generation.

[0093] In this embodiment, during the physical change process simulation phase, the image generation system discretizes the simulation process according to a preset time step. The time step is determined based on the process duration of potential abnormal links and the simulation accuracy requirements. The shorter the process duration, the smaller the time step, allowing the details of the simulation process to be accurately captured. At each time step, the physical field coupled simulation model calculates the morphological parameters (such as line width, pad height, component position, etc.), material performance parameters (such as resistivity, thermal conductivity, mechanical strength, etc.), and surrounding environmental parameters (such as temperature, humidity, pressure, etc.) of the failed circuit board. The image generation system converts the above parameter data into pixel raw data that meets the image generation requirements. The conversion method is to map the morphological parameters to the spatial position and shape of the image, the material performance parameters to the grayscale or color values ​​of the image, and the surrounding environmental parameters to the background features of the image. For example, areas with higher temperatures correspond to brighter colors, and areas with higher stress correspond to darker grayscale, so that the pixel raw data can accurately reflect the state of the circuit board at that time step.

[0094] Step 1627: Based on the pixel distribution rules and color mapping standards in the image generation rules, update the original pixel data of each time step to generate the corresponding single-frame image. Use a time-series stitching algorithm to stitch the single-frame images according to the time sequence of physical changes to generate an initial intermediate state image sequence. Extract the image frame corresponding to the key process node of the potential abnormal link from the initial intermediate state image sequence as the initial intermediate state image.

[0095] In this embodiment, the image generation system updates the original pixel data at each time step based on the pixel distribution rules and color mapping standards in the image generation rules. The pixel distribution rules are used to adjust the pixel density of different feature regions so that the pixel density of the feature regions corresponding to the abnormal variables meets the requirements. The color mapping standards are used to adjust the color distribution of the image so that different types of abnormal variables correspond to different colors, and the color saturation is positively correlated with the correlation strength. After the update processing is completed, the original pixel data at each time step is converted into a corresponding single-frame image, and the image format is a lossless bitmap containing RGB three channels. Further, the image generation system stitches the single-frame images according to the time sequence of physical changes using a temporal stitching algorithm to generate an initial intermediate state image sequence. The sequence contains single-frame images corresponding to all time steps from the start to the end of the potential abnormal link. Finally, the image generation system extracts image frames corresponding to the key process nodes of the potential abnormal link from the initial intermediate state image sequence as initial intermediate state images. The key process nodes include the process start time, parameter mutation time, process end time, etc., so that the extracted images can reflect the key states that lead to failure in the potential abnormal link.

[0096] Step 163: Optimize the initial intermediate state image based on the image generation rules, supplement the transient feature information in the initial intermediate state image with the transient scene reproduction algorithm, and adjust the temporal parameters of image generation using the logic transformation unit so that the generated image matches the temporal dimension features of the potential abnormal link, thereby obtaining the optimized intermediate state image.

[0097] Step 1631: Analyze the resolution standard, pixel arrangement specification and feature visualization mapping relationship in the image generation rules, extract the target parameters for initial intermediate state image optimization, and construct an image optimization parameter set containing spatial optimization parameters, temporal optimization parameters and feature supplementation parameters based on the target parameters.

[0098] In this embodiment, the image generation system parses the resolution standard, pixel arrangement specification, and feature visualization mapping relationship in the image generation rules. The resolution standard includes the image's width, height, and pixel density requirements; the pixel arrangement specification includes the pixel arrangement method, region division rules, and pixel positions corresponding to feature points; the feature visualization mapping relationship includes the visualization correspondence rules between failed features and abnormal variables, color mapping standards, and the correspondence between feature intensity and pixel brightness. Then, the image generation system extracts target parameters for optimizing the initial intermediate state image. The target parameters include the target resolution, target pixel arrangement method, and target feature mapping rules. Based on the above target parameters, the image generation system constructs an image optimization parameter set that includes spatial optimization parameters, temporal optimization parameters, and feature supplementation parameters. The spatial optimization parameters include resolution adjustment coefficients, pixel rearrangement rules, and region completion rules; the temporal optimization parameters include frame interval adjustment coefficients, temporal coherence constraints, and key event temporal markers; and the feature supplementation parameters include transient feature types, feature supplementation positions, and feature intensity mapping rules.

[0099] Step 1632: Based on the spatial optimization parameters in the image optimization parameter set, perform spatial resolution calibration on the initial intermediate state image, use bilinear interpolation algorithm to improve the pixel density of the initial intermediate state image, so that the resolution of the calibrated initial intermediate state image matches the requirements of the image generation rule, and rearrange the image pixels according to the pixel arrangement specification.

[0100] In this embodiment, the image generation system performs spatial resolution calibration on the initial intermediate state image based on the spatial optimization parameters in the image optimization parameter set. A bilinear interpolation algorithm is used to increase the pixel density of the initial intermediate state image. The algorithm generates new pixels by calculating the weighted average of four adjacent pixels, with the weights determined based on the distance between the new pixel and its neighbors, ensuring that the enhanced image maintains good sharpness and detail. Then, the image generation system compares the calibrated image resolution with the requirements of the image generation rules. If the requirements are not met, the interpolation process is repeated until the resolution is perfectly matched. Furthermore, the image generation system rearranges the image pixels according to a pixel arrangement specification. This specification includes the pixel positions of different functional areas, the pixel coordinates of feature points, and the pixel distribution in areas corresponding to abnormal variables. For example, the position of the failure point is adjusted to the center of the image, and the pixels in areas corresponding to abnormal variables are arranged according to a rule that the correlation strength gradually decreases from the center to the periphery.

[0101] Step 1633: Activate the preset transient scene reproduction algorithm to extract key transient process events based on the production process timing record data of the potential abnormal links. The key transient process events include the process states corresponding to the moment of material input, the moment of equipment start-up and shutdown, and the moment of sudden change in process parameters.

[0102] In this embodiment, the image generation system activates a preset transient scene reconstruction algorithm to retrieve production process timeline data for potential abnormal processes. This data includes all time points, parameter change records, and equipment operation records during the process. Then, the algorithm extracts key transient process events based on this data. These key transient process events include the process states corresponding to the instants of material input (e.g., solder wire input, ink spraying), equipment start-up and shutdown (e.g., pick-and-place machine start-up, reflow oven shutdown), and sudden changes in process parameters (e.g., sudden increase in soldering temperature, sudden change in printing pressure). Each key transient process event includes the event occurrence time, event type, corresponding parameter change value, and the type of equipment involved.

[0103] Step 1634: For each critical transient process event, combine the transient fluctuation data in the abnormal variable data and the transient correlation features in the failure features, and generate the corresponding transient feature image block through the feature enhancement generation algorithm. The transient feature image block includes a transient current distribution image block, a transient temperature gradient image block and a transient material distribution image block.

[0104] In this embodiment, for each key transient process event, the image generation system retrieves transient fluctuation data from the abnormal variable data and transient correlation features from the failure features. The transient fluctuation data includes the fluctuation amplitude of process parameters, equipment parameter deviations, and changes in environmental variables at the time of the event. The transient correlation features include the morphological changes, correlation strength changes, and conduction path changes of the corresponding failure features at the time of the event. Then, the image generation system generates corresponding transient feature image blocks through a feature enhancement generation algorithm. Based on the transient fluctuation data and transient correlation features, the algorithm generates image blocks containing transient current distribution, transient temperature gradient, and transient material distribution. For example, the instant of material input corresponds to a transient material distribution image block, the instant of equipment start-up and shutdown corresponds to a transient current distribution image block, and the instant of sudden change in process parameters corresponds to a transient temperature gradient image block. The resolution of each transient feature image block is consistent with the spatially optimized initial intermediate state image, and the color mapping follows the standard in the image generation rules to achieve seamless integration with the initial intermediate state image.

[0105] Step 1635: The transient feature image block is fused with the spatially optimized initial intermediate state image, and a weighted average fusion algorithm is used to smooth the transition of pixels in the overlapping areas of the images to obtain the fused image.

[0106] In this embodiment, the image generation system determines the corresponding position of the transient feature image block in the spatially optimized initial intermediate state image. The position is determined based on the image region corresponding to the moment the key transient process event occurs, such as the solder application area in the image at the moment of material delivery, and the equipment action area in the image at the moment of equipment start-up and shutdown. The image generation system fuses the transient feature image block with the initial intermediate state image, using a weighted average fusion algorithm to smooth the transition of pixels in the overlapping areas of the images. The weighting coefficient is determined based on the correlation strength between the image block and the initial image; the higher the correlation strength, the higher the weight of the corresponding image block. This ensures that the fused image retains the overall state of the initial image while accurately presenting the transient feature information. After the fusion process is completed, a fused image is generated. The transient feature region in the image transitions naturally with the surrounding area, without obvious splicing marks, and accurately reflects the circuit board state corresponding to the key transient process event.

[0107] Step 1636: Call the logic transformation unit to extract the time dimension feature parameters of the potential abnormal link. The time dimension feature parameters include process duration, parameter change cycle and key event trigger sequence.

[0108] In this embodiment, the image generation system calls the logic conversion unit to extract the time-dimensional feature parameters of potential abnormal links. The unit retrieves the production process standard data and actual production time sequence record data of the potential abnormal links. The process standard data includes requirements for process duration, parameter change cycle, and key event triggering sequence requirements. The actual production time sequence record data includes the actual process duration, actual parameter change cycle, and actual key event triggering time. Then, the unit extracts time-dimensional feature parameters based on the above data. The process duration is the deviation between the actual process duration and the standard duration; the parameter change cycle is the ratio of the actual parameter change cycle to the standard cycle; and the key event triggering sequence is the time difference between the actual triggering time and the standard time. This allows the parameters to reflect the abnormality of the potential abnormal links in the time dimension. All time-dimensional feature parameters are standardized and fall within the range of 0 to 1, with higher values ​​indicating a more severe abnormality.

[0109] Step 1637: Construct a temporal adjustment model based on the time dimension feature parameters, input the fused image into the temporal adjustment model to optimize the frame interval duration and temporal coherence, and obtain the temporally adjusted image.

[0110] In this embodiment, the image generation system constructs a timing adjustment model based on time-dimensional feature parameters. The model adopts a recurrent neural network architecture, comprising three core modules: an input layer, a timing processing layer, and an output layer. The input layer contains the fused image and time-dimensional feature parameters. The timing processing layer mines the temporal correlation between image frames through gated recurrent units. The output layer is the adjusted image frame sequence. The image generation system inputs the fused image into the input layer of the timing adjustment model. The input layer converts the image into a high-dimensional feature vector and concatenates the time-dimensional feature parameters with the feature vector. The timing processing layer adjusts the frame interval of the image frames based on the process duration in the parameters. The larger the deviation in process duration, the smaller the frame interval, enabling a more detailed representation of abnormal process processes. It also adjusts the temporal coherence of the image frames based on the parameter change period in the parameters. The larger the ratio of parameter change periods, the smoother the feature changes between frames, ensuring that the timing changes conform to the actual process rules. Finally, it adjusts the position of the image frames corresponding to key events based on the key event trigger timing in the parameters, ensuring that the trigger time is consistent with the actual production timing. Finally, the output layer outputs the time-adjusted image. The frame interval, temporal continuity, and key event positions of the image frame sequence all meet the requirements of the time dimension feature parameters, which can accurately reflect the time dimension features of potential abnormal links.

[0111] Step 1638: Use a feature matching verification algorithm to correlate and verify the time-adjusted image with the failure features and the abnormal variable data to obtain the matching degree between the features presented by the time-adjusted image and the actual process features. If the matching degree does not reach the preset matching threshold, return to readjust the time-series parameters and supplement the transient features until the matching degree reaches the preset matching threshold to obtain the optimized intermediate state image.

[0112] In this embodiment, the image generation system uses a feature matching verification algorithm to correlate and verify the time-adjusted image with failure features and abnormal variable data. The algorithm extracts key features from the time-adjusted image, including the failure point morphology, the corresponding region of the abnormal variable, and the transient feature region. These features are then matched with corresponding features in the failure feature set and corresponding parameters in the abnormal variable data to calculate the feature similarity and parameter matching degree. The similarity is determined based on the cosine similarity of the feature vectors, and the matching degree is determined based on the parameter deviation value. The image generation system weights and fuses the feature similarity and parameter matching degree to obtain the overall matching degree. The weighting coefficient is determined based on weight parameters trained on historical data. The overall matching degree is between 0 and 1, with higher values ​​indicating better matching. If the overall matching degree does not reach the preset matching threshold, the system returns to step 1637 to readjust the time-series parameters and returns to step 1634 to replenish the transient feature content until the matching degree reaches the preset matching threshold. The resulting image is the optimized intermediate state image, which accurately reflects the state of the failed circuit board in the potential abnormal stage and highly matches the failure features and abnormal variable data.

[0113] Step 164: Associate and match the optimized intermediate state image with the failure features and the abnormal variable data. After confirming that the matching is qualified, obtain the intermediate state image of the failed circuit board corresponding to the potential abnormal link. The intermediate state image is used to present the morphological state and variable distribution state of the failed circuit board in the potential abnormal link.

[0114] In this embodiment, the image generation system correlates and matches the optimized intermediate state image with failure features and abnormal variable data. A feature extraction algorithm extracts all visual features from the optimized intermediate state image, including the location and shape of the failure point, the state of surrounding circuitry, the color distribution of the area corresponding to the abnormal variable, and details of the transient feature area. Further, the system matches these visual features with corresponding features in the failure feature set, and simultaneously matches the variable distribution features in the image with the abnormal variable data, calculating the matching degree of each feature and the overall matching degree. If the overall matching degree reaches a preset acceptable threshold, the matching is considered successful, and the image is used as the intermediate state image corresponding to the potentially abnormal stage of the circuit board. The image includes the morphological state of the circuit board (e.g., circuit breakage, pad oxidation, component misalignment) and the variable distribution state (e.g., temperature distribution, stress distribution, material distribution), providing intuitive visual evidence for failure tracing analysis and process optimization. If the overall matching degree does not reach the acceptable threshold, the system returns to step 163 for re-optimization until the matching degree is acceptable.

[0115] Optionally, after steps 110-160, this embodiment further includes:

[0116] Step 210: Based on the intermediate state image and the failure features, perform process deviation transmission logic mining. Input the intermediate state image into the feature association analysis model, and mine the visual association features related to the process deviation in the intermediate state image through the image feature parsing unit of the feature association analysis model. Then, fuse the visual association features with the failure features to obtain the process deviation-failure feature association set.

[0117] In this embodiment, the image generation system inputs the intermediate state image into a feature association analysis model. This model adopts an architecture combining convolutional neural networks and graph neural networks, comprising three core modules: an image feature parsing unit, a feature fusion unit, and a logic mining unit. The image feature parsing unit extracts visual features from the intermediate state image through multi-scale convolutional layers, including morphological features, color features, texture features, and spatial distribution features. It then identifies visual association features related to process deviations using an association rule mining algorithm. These visual association features include morphological changes, abnormal color distributions, and texture abrupt changes corresponding to process deviations, such as discoloration areas caused by excessive temperature or deformation areas caused by excessive stress. Next, the feature fusion unit fuses the visual association features with the failure features, merging their feature vectors along the feature dimension to obtain a process deviation-failure feature association set. Each association entry includes a visual association feature identifier, a failure feature identifier, and association strength, with the association strength determined based on the similarity of the feature vectors.

[0118] Step 220: Logically complete the process deviation-failure feature association set by combining the abnormal variable data of the potential abnormal links, determine the transmission path of process deviation from the initial state to the failure by using the time sequence association analysis algorithm, and generate the process deviation transmission logic chain.

[0119] In this embodiment, the image generation system logically completes the process deviation-failure feature association set by combining abnormal variable data of potential abnormal links. It matches process fluctuation data and equipment operating parameter deviation data in the abnormal variable data with visual association features in the association set, supplementing the initial triggering factors of the process deviation, such as the initial deformation region in the visual association features corresponding to equipment operating parameter deviations. The image generation system determines the transmission path of the process deviation from its initial state to causing failure through a temporal correlation analysis algorithm. Based on the temporal information of the intermediate state image sequence, the algorithm sorts out the order of appearance of visual association features, such as the initial temperature deviation region leading to the line deformation region, and the line deformation region leading to the line breakage failure feature. A causal inference algorithm verifies the causal relationship of each link, excluding links that only have temporal correlation but no causal relationship. Finally, all links with causal relationships are linked together in temporal order to generate a process deviation transmission logic chain. The logic chain includes triggering factors, intermediate transmission links, failure features, the correlation strength of each link, and the transmission time, clearly presenting the complete process of how the process deviation causes failure.

[0120] Step 230: Based on the process deviation transmission logic chain, reverse the derivation of the process parameter adjustment logic, compare the process deviation transmission logic chain with the preset process standard logic library, discover the process parameter correction direction corresponding to the key nodes of deviation transmission, and generate a process optimization logic diagram.

[0121] In this embodiment, the image generation system reverse-engineers the process parameter adjustment logic based on the process deviation transmission logic chain. It traces back from the failure characteristic link of the logic chain to the initial triggering factor, determining the process parameters corresponding to each transmission link. For example, the welding temperature parameter corresponds to the circuit deformation link, and the stress control parameter corresponds to the circuit fracture link. Then, the image generation system compares the process deviation transmission logic chain with a preset process standard logic library. The process standard logic library contains standard process parameters, parameter adjustment rules, deviation correction methods, etc., for each production link. The comparison includes the deviation values ​​between the actual process parameters and the standard parameters, the differences between the transmission links and the standard logic, and the differences between the failure characteristics and the standard output. The image generation system identifies the process parameter correction direction corresponding to the key nodes of deviation transmission. The key nodes are the transmission links that have the greatest impact on the failure characteristics. The correction direction is determined based on the sign of the deviation value and the requirements of the standard logic. For example, if the actual temperature is higher than the standard temperature, the correction direction is to reduce the welding temperature; if the actual stress is higher than the standard stress, the correction direction is to reduce the equipment pressure. Finally, the image generation system integrates the correction direction, key nodes, and relationships into a process optimization logic diagram. This diagram is presented in a visual topology structure and includes process parameter nodes, transmission paths, correction directions, optimization targets, etc., which can intuitively guide the process optimization work.

[0122] Step 240: Associate and map the process optimization logic diagram with the visual features of process deviations in the intermediate state image to generate process optimization suggestions that include the correspondence between parameter adjustment logic and visual features. Guide the dynamic calibration of process parameters in potential abnormal links through the process optimization suggestions.

[0123] In this embodiment, the image generation system correlates and maps the process optimization logic diagram with the visual features of process deviations in intermediate state images. It determines the visual feature regions in the intermediate state images corresponding to each key node in the process optimization logic diagram, such as the discoloration region of the circuit corresponding to excessive temperature or the deformation region of the circuit corresponding to excessive pressure. Furthermore, it binds the process parameter correction direction of each key node to the corresponding visual feature region, generating a correspondence between parameter adjustment logic and visual features. This correspondence includes visual feature region identifiers, key node identifiers, parameter correction directions, and suggested correction magnitudes. The image generation system integrates this correspondence into process optimization suggestions, which include potential abnormal link identifiers, process optimization logic diagrams, a table of parameter adjustment and visual feature correspondences, and optimization implementation steps. This guides operators to quickly locate process deviations and adjust parameters. Finally, the image generation system sends the process optimization suggestions to the production control system. Based on the suggestions, the production control system dynamically calibrates the process parameters of potential abnormal links, adjusting equipment operating parameters, environmental parameters, and material supply parameters in real time. This ensures that the process status in subsequent production processes meets standard requirements, preventing similar failures from recurring.

[0124] Optionally, after steps 110-160, this embodiment further includes:

[0125] Step 310: Align the intermediate state image with the full-process time-series data of the production chain, establish the time-series correspondence between the intermediate state image and the operation data of each production link through a timestamp matching algorithm, and generate a time-series associated image dataset.

[0126] In this embodiment, the image generation system retrieves full-process time-series data of the production chain. This data includes the start and end times, key event trigger times, and process parameter time-series records for each production stage, with all data containing a unique timestamp. The image generation system uses a timestamp matching algorithm to align intermediate state images with the full-process time-series data. The algorithm extracts the time range of potential abnormal stages corresponding to the intermediate state images and further matches intermediate state image frames within that time range with the time-series data of the corresponding production stages. Each image frame corresponds to operational data within a timestamp range. After matching, a time-series associated image dataset is generated. This dataset contains the identifier of each image frame, its corresponding timestamp, the identifier of the corresponding production stage, and the corresponding operational data storage address, enabling rapid association queries between images and production data, providing a data foundation for subsequent source tracing evidence mining.

[0127] Step 320: Based on the time-series associated image dataset and the failure features, perform source tracing evidence association mining. Input the time-series associated image dataset into the evidence feature extraction model, mine the core evidence features representing production anomalies in the intermediate state images, and logically associate the core evidence features with the failure features to obtain evidence-failure association logical features.

[0128] In this embodiment, the image generation system inputs a time-series correlated image dataset into an evidence feature extraction model. This model employs a convolutional neural network architecture, comprising three core modules: a feature extraction layer, an evidence recognition layer, and an output layer. The feature extraction layer extracts high-dimensional features from the image through multi-scale convolution. The evidence recognition layer identifies core evidence features characterizing production anomalies using a pre-trained classifier. These core evidence features include morphological changes corresponding to process deviations, color distributions corresponding to abnormal variables, and texture mutations corresponding to transient events. Then, the image generation system logically correlates the core evidence features with failure features, verifying the causal relationship between them using a causal inference algorithm. Feature combinations that are only correlated but not causal are excluded, resulting in evidence-failure correlation logical features. Each correlation logical feature includes a core evidence feature identifier, a failure feature identifier, causal correlation strength, and a transmission path, clearly presenting the intrinsic connection between the evidence features and the failure features.

[0129] Step 330: By using the correlation logic verification algorithm, the abnormal variable data of the potential abnormal links are integrated to strengthen the evidence-failure correlation logic features, and a strengthened evidence correlation feature set is generated.

[0130] In this embodiment, the image generation system strengthens the evidence-failure correlation logic features by fusing anomalous variable data from potential abnormal links using an association logic verification algorithm. The algorithm matches process fluctuation data and equipment operating parameter deviation data from the anomalous variable data with the core evidence features in the correlation logic features to verify the existence of the anomalous variables corresponding to the evidence features. If they exist, the strength of the correlation logic feature is enhanced. Furthermore, the system matches the changing trends of the anomalous variable data with the occurrence sequence of the failure features to verify their temporal consistency. If they are consistent, the strength of the correlation logic feature is further enhanced. The image generation system integrates the enhanced correlation logic features into a strengthened evidence correlation feature set. Each feature set entry includes a core evidence feature identifier, a failure feature identifier, the enhanced causal correlation strength, the anomalous variable matching status, and the temporal consistency verification result, ensuring that each correlation logic feature is supported by anomalous variable data, thus increasing its credibility.

[0131] Step 340: Determine the failure tracing evidence chain based on the enhanced evidence association feature set, and integrate the core evidence features, failure features and abnormal variable data hierarchically according to the temporal association relationship and logical association strength to obtain the evidence chain characterizing the failure tracing process.

[0132] In this embodiment, the image generation system determines the failure tracing evidence chain based on a strengthened evidence association feature set. Using failure features as the core, it concatenates corresponding core evidence features and anomalous variable data according to temporal correlations. These temporal correlations are determined based on timestamps in the temporally correlated image dataset, ensuring the correct temporal order of the evidence chain. Furthermore, the image generation system hierarchically integrates the evidence chain according to the strength of logical association. Evidence with higher correlation strength is located at a more core level; for example, the core evidence feature with the strongest causal correlation is located at the first level, the corresponding anomalous variable data at the second level, and secondary evidence features at the third level. Finally, the image generation system integrates evidence from all levels into an evidence chain representing the failure tracing process. This evidence chain includes failure feature identifiers, a list of core evidence features, a list of anomalous variable data, temporal correlations, logical correlation strength levels, and tracing conclusions, providing sufficient and credible evidence support for liability determination, process improvement, and quality control in failure issues.

[0133] It should be understood that the key steps in this embodiment (such as anomaly correlation analysis and feature mining) need to handle multiple parameters with different physical meanings and dimensions. Directly using the original numerical values ​​will result in invalid mathematical operations or meaningless results due to unit mismatch. Adaptive normalization, as a conventional and mature technology in this field, can effectively solve this problem. Technicians can select methods such as minimum-maximum normalization according to the parameter characteristics to map multi-source parameters to a unified dimensionless scale, eliminate unit differences, and retain the relative importance in process correlation. This can be completed without creative labor. This processing runs through the entire process, ensuring the feasibility of the solution: achieving scale unification between image features and physical parameters in the failure feature mining stage; improving the accuracy of node embedding and inference in anomaly correlation analysis; facilitating the conversion into unitless generation instructions when determining generation rules; and ensuring the smooth conversion of data into original pixel data in physical process simulation. In this way, all calculation and fusion steps can be mathematically coordinated, supporting the successful execution of the source tracing task from failure location to intermediate state image generation.

[0134] The reverse image generation method for PCB failure tracing provided in this application constructs a complete technical chain from the global logical level, deducing production anomalies from failure results and reconstructing intermediate states: through the collaborative processing of pre-defined failure location benchmark region division, failure point morphology detail acquisition, and layout location correlation information mining, it breaks through the limitation of traditional failure analysis focusing only on isolated failure phenomena, realizing the targeted identification of failure features and global correlation logic mining, and can accurately capture the intrinsic relationship between the micro-morphology of failure points and the PCB layout; based on the failure features, the production process chain anomaly correlation analysis establishes a value by quantifying the correlation degree between each production process and the failure features. The implicit correlation mapping mechanism between failure features and production anomalies solves the problem of accurately locating the cause from the result in reverse tracing, effectively anchoring the source production link that triggers the failure. By determining image generation rules through the correlation and fusion of abnormal variable data and failure features, and combining potential abnormal link scene labels to achieve reverse state generation, a scene restoration system based on production logic is constructed. This system can transform abstract production abnormal variables into visualized intermediate state images, overcoming the limitation of forward image generation that only focuses on production output. It allows abnormal states hidden in the production process to be presented intuitively, providing traceable and verifiable visualized evidence support for circuit board failure tracing. Thus, this embodiment achieves end-to-end collaboration between failure feature identification, abnormal link locking, and intermediate state image generation, improving the accuracy, efficiency, and intuitiveness of circuit board failure tracing.

[0135] Please see Figure 2The figure is a schematic diagram of the basic structure of an image generation system 200 provided in an embodiment of this application. The image generation system 200 includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the image generation methods applied to circuit board production.

[0136] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0137] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of the image generation system reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the image generation system to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0138] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. An image generation method for use in circuit board production, characterized in that, The method includes: In response to a tracing task request for a failed circuit board, the failed circuit board is divided into regions based on a preset failure location benchmark to locate the failure area of ​​the failed circuit board. Collect morphological detail images of the failure points within the failure area, and determine the layout location association information of the failure points in the failure circuit board through circuit board layout mapping processing; Based on the layout location association information, feature mining processing is performed on the morphological detail image to obtain the failure features corresponding to the morphological detail image; An anomaly correlation analysis is performed on the production chain of the failed circuit board based on the failure characteristics to obtain an anomaly tendency analysis result that includes the correlation degree value between each production link and the failure characteristics. The process involves identifying potential abnormal links in the production chain based on the anomaly tendency analysis results and extracting corresponding abnormal variable data. Image generation rules are then determined based on the correlation and fusion processing results of the abnormal variable data and the failure characteristics. This includes: extracting abnormal variable data from the potential abnormal links based on a preset variable extraction template. The abnormal variable data includes process fluctuation data, equipment operating parameter deviation data, environmental variable change data, and material characteristic deviation data. The extracted abnormal variable data is then standardized to obtain standardized abnormal variable data. The dynamic correlation between the standardized abnormal variable data and the failure characteristics is then mined to determine the causal relationship between them. The dynamic correlation and the causal relationship are then transformed into spatial transformation parameters, pixel distribution rules, and color mapping standards required for image generation. These parameters, pixel distribution rules, and color mapping standards are then optimized using preset visualization rules to generate the image generation rules. The image generation rules include the resolution of the generated image, pixel arrangement, and feature visualization mapping relationships. Based on the scene label of the potential abnormal link, the state of the failed circuit board in the potential abnormal link is reverse generated according to the image generation rule to obtain the intermediate state image of the failed circuit board in the potential abnormal link.

2. The method as described in claim 1, characterized in that, The step of performing feature mining processing on the morphological detail image based on the layout position association information to obtain the failure features corresponding to the morphological detail image includes: Based on the layout position association information, the effective analysis area of ​​the morphological detail image is determined, a multi-scale fusion analysis algorithm is invoked, and the morphological detail image is input into the multi-scale fusion analysis algorithm. Initial features of the morphological detail image at the pixel level, texture level, and structural level are extracted by feature extraction units at different scales. A hierarchical association index is established based on the layout location association information. The initial features at each scale are matched and associated with the hierarchical association index. Candidate features associated with the layout location of the failure point are selected by a directional identification algorithm. The association logic mining model is invoked, and the candidate features and their corresponding layout position association information are input into the association logic mining model. The association logic mining model is used to mine the association logic topology between the candidate features, and the candidate features with association relationships are fused to obtain fused features. Based on the aforementioned association logic topology, a feature association confidence evaluation strategy is generated to quantitatively evaluate the association logic confidence of the fused features and generate feature association confidence parameters. A dynamic weight allocation model is created by combining the layout location association information. The feature association confidence parameter is used as input, and dynamic weights are assigned to information of different association dimensions in the fused features through the dynamic weight allocation model. Based on the dynamic weights, the fusion features are subjected to hierarchical enhancement processing, and a dynamic mapping relationship between the fusion features and the layout position association information is established. Based on the dynamic mapping relationship, the fusion features after layered enhancement are associated and integrated to obtain the failure features corresponding to the morphological detail image. The failure features contain dual attributes of morphological features and layout association features, and each failure feature carries a corresponding layout position association marker and feature association confidence parameter.

3. The method as described in claim 1, characterized in that, The step of performing anomaly correlation analysis on the production chain of the failed circuit board based on the failure characteristics, and obtaining anomaly tendency analysis results including the correlation degree values ​​between each production link and the failure characteristics, includes: The failure features are structured and encoded to obtain a structured failure feature dataset. The attributes of each production link in the production chain are decomposed, and the process parameters, operating procedures and material information of each production link are extracted to generate a production link attribute dataset. A latent association graph is generated based on the structured failure feature dataset and the production process attribute dataset. The failure features and each production process are respectively used as nodes of the latent association graph, and the potential association relationships between the failure features and each production process are used as edges of the latent association graph. The implicit association graph is input into a preset graph neural network model. The node embedding layer of the graph neural network model is used to perform feature embedding processing on each node. The multi-hop inference algorithm is used to mine the multi-order association relationship between the failure feature node and the production link node in the implicit association graph. Based on the aforementioned multi-level correlation, the causal correlation strength between each production link and failure feature is analyzed using a causal inference algorithm. The collaborative anomaly patterns corresponding to different combinations of production links and failure features are identified, and the correlation strength between each production link and failure feature is quantified to obtain the correlation degree value. By integrating the correlation degree values, collaborative anomaly patterns, and corresponding production process information, an anomaly tendency analysis result containing the correlation degree values ​​between each production process and failure characteristics is obtained.

4. The method as described in claim 3, characterized in that, The process involves inputting the implicit association graph into a preset graph neural network model, performing feature embedding processing on each node through the node embedding layer of the graph neural network model, and using a multi-hop inference algorithm to mine the multi-order association relationships between failed feature nodes and production process nodes in the implicit association graph, including: The latent association graph is preprocessed, including outlier node removal and feature dimension addition, to obtain a preprocessed latent association graph. The preprocessed latent association graph is input into a preset graph neural network model. The node embedding layer of the graph neural network model is called, and a node embedding algorithm based on the attention mechanism is used to encode the features of each node. The inherent attribute features of each node are fused with the graph topology features to generate an initial node embedding vector. Based on the initial node embedding vector, a node feature similarity matrix is ​​constructed. The feature matching degree between each node is solved by the cosine similarity calculation method. The feature matching degree is used as a constraint condition for node embedding optimization. The initial node embedding vector is iteratively optimized until the difference between the node embedding vectors generated by two adjacent iterations is less than a preset difference threshold, and the optimized node embedding vector is obtained. The preset multi-hop inference algorithm is started. Based on the optimized node embedding vector, the inference hop count threshold is set. Each failure feature node is used as an independent starting node. The algorithm is traversed hop by hop according to the edge connection relationship of the preprocessed implicit association graph. The first-order association production link node directly connected to the starting node is mined. Then, the mined first-order association production link node is used as the intermediate transmission node. The algorithm continues to traverse and mined the second-order association node connected to the intermediate transmission node. This process is repeated until the preset inference hop count threshold is reached. During the multi-hop inference process, the correlation of process parameters in the production link attribute dataset and the feature correlation attributes in the structured failure feature dataset are combined to assign dynamic correlation weights to the correlation relationships obtained by each hop mining. The dynamic correlation weights are positively correlated with the feature matching degree between nodes and the tightness of process correlation. After hierarchically labeling the associations corresponding to different hop counts, a multi-level association graph is generated by combining the association nodes, association edges, and corresponding dynamic association weights at each level. Based on the multi-level association graph, the association path information between each failure feature node and the production link node is extracted. Based on the association path information, the length of each association path and the sum of the dynamic association weights of each association edge on the path are quantitatively statistically analyzed, and effective association paths with a sum of dynamic association weights exceeding a preset weight threshold are selected. By combining the effective association paths obtained through screening with the corresponding node association information and weight information, a multi-level association relationship is generated between the failure feature nodes and the production process nodes, including association level, association strength and association path.

5. The method as described in claim 1, characterized in that, The step of identifying potential abnormal links in the production chain based on the abnormality tendency analysis results and extracting the corresponding abnormal variable data for the potential abnormal links, and determining image generation rules based on the correlation and fusion processing results of the abnormal variable data and the failure features, includes: Extract the correlation degree value from the abnormal tendency analysis results, and identify the production links with a correlation degree exceeding the preset correlation degree threshold as potential abnormal links in the production link chain.

6. The method as described in claim 1, characterized in that, The step of performing reverse generation processing on the state of the failed circuit board at the potential abnormal link according to the scene label of the potential abnormal link and the image generation rule to obtain the intermediate state image of the failed circuit board at the potential abnormal link includes: Based on the scene label of the potential abnormal link, the corresponding scene restoration model is called, the image generation rule is transformed into the input parameters of the scene restoration model, and the abnormal variable data of the potential abnormal link and the failure characteristics are simultaneously input into the scene restoration model. Based on the current state image of the failed circuit board, and combined with the input parameters, the physical process simulation unit is activated to simulate the physical change process of the failed circuit board in the potential abnormal link, and generate an initial intermediate state image. Based on the image generation rule, the initial intermediate state image is optimized, the transient feature information in the initial intermediate state image is supplemented, and the temporal parameters of image generation are adjusted so that the generated image matches the temporal dimension features of the potential abnormal link, thus obtaining the optimized intermediate state image. The optimized intermediate state image is associated and matched with the failure features and the abnormal variable data. After confirming that the matching is qualified, the intermediate state image of the failed circuit board corresponding to the potential abnormal link is obtained. The intermediate state image is used to present the morphological state and variable distribution state of the failed circuit board in the potential abnormal link.

7. The method as described in claim 6, characterized in that, The step of using the current state image of the failed circuit board as a reference, combined with the input parameters, to activate the physical process simulation unit, simulate the physical changes of the failed circuit board at the potential abnormal stage, and generate an initial intermediate state image, including: The input parameters transformed by the image generation rules are decomposed into spatial configuration parameters, physical simulation parameters, and feature mapping parameters, and a one-to-one mapping relationship between the feature points of the standardized current state image and each input parameter is established. Based on the one-to-one mapping relationship, the physical process simulation unit is started, and a physical field coupling simulation model is constructed in combination with the process type of the potential abnormal link. The physical field coupling simulation model is used to simulate the synergistic effect of temperature field, stress field and material action field on the failed circuit board. The process fluctuation data and equipment operating parameter deviation data in the abnormal variable data are transformed into boundary constraints of the physical field coupling simulation model to simulate the dynamic physical environment change process of the failed circuit board in the potential abnormal link. At the same time, combined with the morphological details in the failure characteristics, the deformation, oxidation and circuit connection relationship change process of the circuit board material under the action of the dynamic physical environment are simulated. During the physical change process simulation stage, the simulation process is discretized according to the preset time step. The morphological parameters, material performance parameters and surrounding environment parameters of the failed circuit board are recorded in real time at each time step. The parameter data corresponding to each time step is converted into pixel raw data that meets the requirements of image generation. Based on the pixel distribution rules and color mapping standards in the image generation rules, the original pixel data at each time step is updated to generate corresponding single-frame images. The single-frame images are then stitched together according to the time sequence of physical changes using a time-series stitching algorithm to generate an initial intermediate state image sequence. Image frames corresponding to the key process nodes of the potential abnormal links are extracted from the initial intermediate state image sequence as the initial intermediate state images.

8. The method as described in claim 6, characterized in that, The optimization of the initial intermediate state image based on the image generation rule, supplementing the transient feature information in the initial intermediate state image, and adjusting the temporal parameters of image generation to make the generated image match the temporal dimension features of the potential abnormal link, to obtain the optimized intermediate state image, includes: The resolution standard, pixel arrangement specification and feature visualization mapping relationship in the image generation rule are analyzed to extract the target parameters for the initial intermediate state image optimization. Based on the target parameters, an image optimization parameter set including spatial optimization parameters, temporal optimization parameters and feature supplementation parameters is constructed. Based on the spatial optimization parameters in the image optimization parameter set, the initial intermediate state image is spatially calibrated, and the pixel density of the initial intermediate state image is improved by using a bilinear interpolation algorithm, so that the resolution of the calibrated initial intermediate state image matches the requirements of the image generation rules, and the image pixels are rearranged according to the pixel arrangement specification. Based on the production process timing record data of the potential abnormal links, key transient process events are extracted. The key transient process events include the process states corresponding to the moment of material input, the moment of equipment start-up and shutdown, and the moment of sudden change in process parameters. For each critical transient process event, a corresponding transient feature image block is generated by combining the transient fluctuation data in the abnormal variable data and the transient correlation features in the failure features through a feature enhancement generation algorithm. The transient feature image block includes a transient current distribution image block, a transient temperature gradient image block, and a transient material distribution image block. The transient feature image block is fused with the spatially optimized initial intermediate state image, and a weighted average fusion algorithm is used to smooth the transition of pixels in the overlapping areas of the images to obtain the fused image. Extract the time-dimensional feature parameters of the potential abnormal links. The time-dimensional feature parameters include process duration, parameter change cycle and key event triggering sequence. A temporal adjustment model is constructed based on the time dimension feature parameters. The fused image is then input into the temporal adjustment model to optimize the frame interval duration and temporal coherence, resulting in a temporally adjusted image. The time-adjusted image is correlated with the failure features and the abnormal variable data by a feature matching verification algorithm to obtain the matching degree between the features presented by the time-adjusted image and the actual process features. If the matching degree does not reach the preset matching threshold, the time parameters and transient feature supplements are readjusted until the matching degree reaches the preset matching threshold to obtain the optimized intermediate state image.

9. An image generation system, characterized in that, include: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; and when the computer program is executed by the processor, causing the processor to implement the image generation method for circuit board production as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the image generation method for circuit board production as described in any one of claims 1-8.

Images