Circuit board quality inspection test data analysis method and system based on deep learning

Abstract

The application provides a circuit board quality inspection test data analysis method and system based on deep learning, and belongs to the technical field of electronic manufacturing quality detection. First, a circuit board quality inspection test data set is obtained, and is mapped and bound with the detection feature dimension of a deep learning quality inspection analysis model to generate a bound association set; then, the bound association set is used to drive the model to dynamically expand the detection feature dimension, and after obtaining the expanded detection feature dimension set, the circuit board quality inspection test data set is subjected to deep feature mining according to the detection feature dimension set, and a deep mining feature set is obtained, the model internal node connection and feature transmission rule are adjusted, an adaptive model internal feature conduction path is generated, and finally, the deep mining feature set is subjected to full-dimension feature analysis according to the adaptive feature conduction path set to generate a circuit board quality inspection test data analysis result. The application can dynamically adapt to different circuit board features, deeply mine data information, and improve detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of electronic manufacturing quality inspection technology, and more specifically, to a method and system for analyzing circuit board quality inspection test data based on deep learning. Background Technology

[0002] In the electronics manufacturing industry, circuit boards (PCBs) are core components, and their quality directly determines the performance and reliability of electronic products. Therefore, rigorous quality inspection of PCBs is crucial. Traditional PCB quality inspection data analysis methods primarily rely on manual experience to set fixed inspection feature dimensions, and then process and analyze the PCB quality inspection data based on these preset dimensions. However, this method has several limitations. On the one hand, due to the complex and diverse manufacturing processes of PCBs, different batches and models of PCBs may have different quality characteristics. Fixed feature dimensions cannot comprehensively cover all possible quality problems, easily leading to missed or false detections. On the other hand, traditional methods lack the ability to deeply mine data, failing to extract representative and crucial feature information from massive amounts of quality inspection data, making it difficult to discover potential quality hazards, and thus affecting the overall control and improvement of PCB quality. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for analyzing circuit board quality inspection test data based on deep learning, the method comprising:

[0004] Obtain a circuit board quality inspection test data set, and map and bind the circuit board quality inspection test data set with the detection feature dimensions of the deep learning quality inspection analysis model to generate a binding association set between the data dimensions of the circuit board quality inspection test data set and the detection feature dimensions of the deep learning quality inspection analysis model;

[0005] Based on the binding association set, the deep learning quality inspection analysis model is driven to perform dynamic expansion processing of the detection feature dimensions, so as to obtain the expanded detection feature dimension set of the deep learning quality inspection analysis model.

[0006] The circuit board quality inspection test data set is processed by deep feature mining using the expanded detection feature dimension set to obtain the deep mining feature set of the circuit board quality inspection test data set.

[0007] Based on the deep mining feature set, the node connection and feature transmission rules inside the deep learning quality inspection analysis model are adjusted to generate a feature transmission path inside the model that is adapted to the deep mining feature set, thereby obtaining the adapted feature transmission path set of the deep learning quality inspection analysis model.

[0008] Based on the adapted feature transmission path set, the deep mining feature set is subjected to full-dimensional feature parsing processing to generate the circuit board quality inspection test data analysis results corresponding to the circuit board quality inspection test data set.

[0009] Furthermore, embodiments of the present invention also provide a circuit board quality inspection and testing data analysis system based on deep learning, comprising:

[0010] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned deep learning-based circuit board quality inspection test data analysis method by executing the machine-executable instructions.

[0011] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the aforementioned deep learning-based circuit board quality inspection test data analysis method.

[0012] Based on the above, by mapping and binding the circuit board quality inspection test data set with the detection feature dimensions of the deep learning quality inspection analysis model, a bound association set is generated. This bound association set drives the model to dynamically expand the detection feature dimensions, enabling the model to flexibly adjust the detection feature dimensions according to the characteristics of the actual circuit board quality inspection test data, adapting to the quality inspection needs of different types of circuit boards, and greatly improving the comprehensiveness and accuracy of the inspection. Using the expanded detection feature dimension set to perform deep feature mining on the circuit board quality inspection test data set, a deep-mined feature set can be extracted from massive amounts of data. Based on the deep-mined feature set, the model's internal node connections and feature transmission rules are adjusted to generate a suitable internal feature transmission path, enabling the model to process and analyze deep-mined features more efficiently, further improving the model's accuracy and efficiency in circuit board quality inspection. Finally, based on the adapted feature transmission path set, full-dimensional feature parsing processing is performed on the deep-mined feature set to generate circuit board quality inspection test data analysis results, effectively improving the quality control level of circuit boards and reducing quality risks. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the execution flow of the circuit board quality inspection test data analysis method based on deep learning provided in an embodiment of the present invention.

[0014] Figure 2 This is a schematic diagram of exemplary hardware and software components of the deep learning-based circuit board quality inspection and test data analysis system provided in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a deep learning-based method for analyzing PCB quality inspection test data, provided in one embodiment of the present invention. The following is a detailed description of this deep learning-based method for analyzing PCB quality inspection test data.

[0016] Step S110: Obtain the circuit board quality inspection test data set, and perform mapping and binding processing on the circuit board quality inspection test data set and the detection feature dimension of the deep learning quality inspection analysis model to generate a binding association set of the data dimension of the circuit board quality inspection test data set and the detection feature dimension of the deep learning quality inspection analysis model.

[0017] This embodiment uses the mass production quality inspection scenario of high-density multilayer circuit boards (HDI) as an application example throughout the text. In this scenario, the circuit board includes 10 circuit layers, a blind and buried via structure, and a 0.1 mm line width and spacing design. The test data covers 28 test indicators in three major categories: electrical performance, physical structure, and material properties. The acquired circuit board quality inspection test data set comes from the intelligent manufacturing execution system (MES). The data format is a mixture of JSON structure and CSV table storage. The JSON data records the coordinate-based defect information of image detection, and the CSV table records the time-series test results of electrical parameters. During the data acquisition process, privacy-sensitive data such as operator employee numbers and equipment serial numbers are processed using a noise addition technique based on differential privacy. By adding random noise conforming to a Laplace distribution to the original data, the processed data meets the needs of statistical analysis but cannot be used to deduce specific individual information.

[0018] Step S111: Extract the inherent data dimensions of each type of test data in the circuit board quality inspection test data set, and generate inherent data dimension description information of the circuit board quality inspection test data set. The inherent data dimension description information of the circuit board quality inspection test data set records the dimension identifier and dimension attribute of each type of test data in the circuit board quality inspection test data set.

[0019] For the aforementioned high-density multilayer circuit board testing data, dimensions were extracted from the original dataset using a data dictionary parsing tool. In the electrical performance data, the inherent dimensions of the conduction resistance testing data include four dimensions: test channel number (character type), excitation current level (discrete numerical type), sampling frequency (continuous numerical type), and three-dimensional coordinates of the test point (composite numerical type). In the physical structure data, the inherent dimensions of the blind via aperture testing data include four dimensions: optical lens magnification (integer type), illumination intensity level (enumerated type), image resolution (integer type), and measurement algorithm version (character type). Dimension identifiers use a three-segment naming rule: "Detection type - physical quantity - dimension number." For example, the dimension identifier for the conduction resistance test channel number is "ER-TC-001," where ER represents electrical resistance testing, TC represents the test channel, and 001 is the dimension number. Dimension attribute records include four metadata items: data type (such as character type, numeric type), value constraints (such as the value range of integer type, the list of optional values ​​for enumeration type), unit of measurement (such as ohm, millimeter), and acquisition accuracy (such as ±0.01), which are stored as a dimension description document in XML format.

[0020] Step S112: Extract the preset detection feature dimensions of the deep learning quality inspection analysis model, and generate the preset detection feature dimension description information of the deep learning quality inspection analysis model. The preset detection feature dimension description information of the deep learning quality inspection analysis model records the dimension identifier and dimension attribute of each detection feature dimension in the deep learning quality inspection analysis model.

[0021] This deep learning-based quality inspection analysis model is fine-tuned based on a pre-trained model from the industrial quality inspection field. The preset detection feature dimensions include three major feature families: Conductivity Reliability Feature Family (CRF), Structural Integrity Feature Family (SIF), and Material Compatibility Feature Family (MCF). The Conductivity Reliability Feature Family contains three core dimensions: Path Resistance Stability Feature (CRF-RES-001), Node Contact Resistance Feature (CRF-CON-002), and Temperature Drift Coefficient Feature (CRF-TEM-003). The Structural Integrity Feature Family contains four core dimensions: Pattern Distortion Rate Feature (SIF-DIS-001), Dielectric Layer Thickness Uniformity Feature (SIF-THI-002), Hole Wall Roughness Feature (SIF-ROU-003), and Metallization Layer Adhesion Feature (SIF-ADH-004). Each preset feature dimension's attributes include feature data type (e.g., vector, matrix), feature extraction algorithm identifier (e.g., Fourier transform, wavelet decomposition), feature importance weight (floating-point number between 0 and 1), and feature fusion priority (integer type), stored as a feature configuration file in ProtocolBuffers format.

[0022] Step S113: Match the inherent data dimension description information of the circuit board quality inspection test data set with the preset detection feature dimension description information of the deep learning quality inspection analysis model to generate an initial matching relationship of dimension representation content.

[0023] Step S1131: The inherent data dimension description information of the circuit board quality inspection test data set is split according to the type of representation content to generate multiple inherent data dimension representation sub-contents, and each inherent data dimension representation sub-content corresponds to a representation attribute of the inherent data dimension.

[0024] Taking the inherent dimensional description information of continuity resistance detection data as an example, it is divided into five sub-contents according to the type of representation content: physical attribute sub-content (describing the electrical characteristics of the resistance), spatial attribute sub-content (describing the location information of the test point), temporal attribute sub-content (describing the temporal characteristics of data acquisition), equipment attribute sub-content (describing the parameter configuration of the testing equipment), and environmental attribute sub-content (describing the temperature and humidity conditions during testing). Each sub-content includes four elements: attribute name, data type definition, value range description, and business meaning description. For example, the physical attribute sub-content includes the "excitation current" attribute, whose data type is defined as "a discrete value ranging from 1 mA to 100 mA," and its business meaning description is "a constant current value applied to the test point for measuring loop resistance." The splitting process adopts an ontology-based semantic parsing method. By constructing an ontology model in the circuit board testing domain, the dimensional description information is matched with the concepts in the ontology model to achieve automatic splitting and classification.

[0025] Step S1132: The preset detection feature dimension description information of the deep learning quality inspection analysis model is split into multiple preset detection feature dimension representation sub-contents according to the same representation content type. The type of the split preset detection feature dimension representation sub-contents is consistent with the representation content type of the inherent data dimension.

[0026] The predefined path resistance stability feature dimension (CRF-RES-001) description information is further divided into five similar characterization sub-contents: physical attribute sub-content (describing the electrical characteristics of resistance stability), spatial attribute sub-content (describing the layout characteristics of resistance test points), temporal attribute sub-content (describing the characteristics of resistance changes over time), device attribute sub-content (describing the algorithm parameters used for feature extraction), and environmental attribute sub-content (describing the coefficients of feature influence by the environment). The physical attribute sub-content includes a "resistance fluctuation threshold" attribute, defined as a "percentage value ranging from 0.01% to 5%", with the business meaning of "an indicator measuring the degree of change in resistance value within the test period". By maintaining the same splitting granularity and type system as the inherent data dimensions, semantic consistency in the subsequent matching process is ensured.

[0027] Step S1133: Compare each inherent data dimension representation sub-content with all preset detection feature dimension representation sub-contents one by one, identify sub-content combinations with the same or similar meanings, and generate similar combinations of representation sub-contents.

[0028] A BERT-based semantic similarity calculation method is used for sub-content comparison, converting the text description of each sub-content into a 768-dimensional semantic vector. Taking the physical attribute sub-content of "excitation current" of conduction resistance and the physical attribute sub-content of "test current intensity" of path resistance stability as examples, the text descriptions of the two sub-contents are first segmented, stop words are removed, and then the text is input into a pre-trained BERT model. The output vector corresponding to the [CLS] tag is extracted as the semantic vector, and then the cosine similarity between the two vectors is calculated. When the similarity value is greater than a preset threshold (0.85), the sub-contents are judged to be a combination of similar meanings. During the comparison process, the uniformity of the units of numerical attributes is checked in advance. For example, the current value in "milliampere" and the current value in "ampere" are uniformly converted to "ampere" before comparison to avoid semantic misjudgment caused by the difference in units.

[0029] Step S1134: Perform similarity representation on each similar combination of representation sub-contents to generate similarity representation information of the corresponding similar combination of representation sub-contents. The similarity representation information of the similar combination of representation sub-contents reflects the degree of similarity and representation association between sub-contents.

[0030] Similarity representation information is stored using a four-tuple structure: (intrinsic sub-content ID, preset sub-content ID, similarity score, association type). The similarity score retains four decimal places, and the association type is divided into three levels: direct matching (01), indirect association (02), and implicit association (03). Direct matching indicates that the physical meanings of the two sub-contents are completely identical, such as the spatial attribute sub-content of "test point coordinates" for conduction resistance and the spatial attribute sub-content of "detection position coordinates" for path resistance stability characteristics. Indirect association indicates that the two sub-contents have a causal relationship, such as the sub-content of excitation current and the sub-content of resistance measurement accuracy. Implicit association indicates that the two sub-contents are linked through a third-party parameter, such as the sub-content of ambient temperature and the sub-content of resistance temperature coefficient. Similarity representation information is visualized and stored in the form of a knowledge graph, where nodes represent sub-contents, edges represent similar combination relationships, and the weight of the edges corresponds to the similarity score.

[0031] Step S1135: Classify all similar combinations of generated representation sub-contents according to their inherent data dimensions, and generate a similar combination subset corresponding to each inherent data dimension. Each similar combination subset corresponding to an inherent data dimension contains all similar combinations of that inherent data dimension and their corresponding similarity representation information.

[0032] The classification is based on the inherent data dimension identifier, grouping all similar combinations belonging to the same dimension identifier into a subset. Taking the test channel number dimension (ER-TC-001) of conduction resistance as an example, its similar combination subset contains three similar combinations: those similar to the "channel identifier" sub-content of the conduction resistance stability feature (similarity 0.92), those similar to the "test port" sub-content of the node contact resistance feature (similarity 0.87), and those similar to the "channel calibration code" sub-content of the temperature drift coefficient feature (similarity 0.76). Each subset is sorted in descending order of similarity score to facilitate subsequent priority selection of high-similarity matches. The classification process is implemented using the GROUPBY operation in the SQL database, aggregating all relevant similar combination records using the inherent data dimension identifier as the grouping key.

[0033] Step S1136: For each subset of similar combinations corresponding to each inherent data dimension, calculate the similarity score of each similar combination within it, select the similar combination with the highest similarity score as the core similar combination of that inherent data dimension, and generate a core similar combination set.

[0034] For each subset of similar combinations, a weighted summation method is used to calculate the comprehensive similarity score, with the following weights: physical attribute sub-content similarity (0.4), spatial attribute sub-content similarity (0.2), temporal attribute sub-content similarity (0.15), equipment attribute sub-content similarity (0.15), and environmental attribute sub-content similarity (0.1). Taking the inherent dimension of blind hole aperture detection data as an example, its similarity combination with the hole wall roughness feature dimension of the structural integrity feature family includes 5 sub-content similarities: physical attribute (0.88), spatial attribute (0.91), temporal attribute (0.72), equipment attribute (0.65), and environmental attribute (0.78), with a comprehensive similarity score of 0.88×0.4+0.91×0.2+0.72×0.15+0.65×0.15+0.78×0.1=0.8175. The similar combination with the highest comprehensive score from each subset is selected as the core similar combination, forming the core similar combination set.

[0035] Step S1137: Assign each core similar combination in the core similar combination set to its corresponding inherent data dimension and preset detection feature dimension to generate assigned core similar combinations. The assigned core similar combinations are associated with the dimension source corresponding to each combination.

[0036] A dimension source identifier is added to each core similarity combination to clarify the complete identifier information of the inherent data dimension and the preset feature dimension in the combination. Taking the core similarity combination of the excitation current level dimension of conduction resistance (ER-CC-002) and the test current intensity sub-content of the path resistance stability feature as an example, the information after attribution binding includes: inherent dimension identifier (ER-CC-002), inherent dimension name (excitation current level), preset feature dimension identifier (CRF-RES-001), preset feature dimension name (path resistance stability feature), similarity score (0.8925), and number of matching sub-contents (5). Attribution binding is achieved through foreign key association in a relational database to establish the association relationship between the core similarity combination table and the inherent dimension table and the preset feature dimension table.

[0037] Step S1138: Arrange all core similarity combinations after they are bound to their respective affiliations according to the degree of association of their representational content to generate an ordered core similarity combination sequence, wherein the ordered core similarity combination sequence reflects the representational association between different dimensions.

[0038] Hierarchical clustering algorithm is used to rank core similarity combinations after they are bound by association, with semantic distance representing content as the clustering basis. First, the semantic distance between any two core similarity combinations is calculated. The distance calculation is based on the word vector space generated by Word2Vec. After converting the text description of the combination into vectors, the Euclidean distance is calculated. Then, agglomerative hierarchical clustering is used to merge the combinations with the closest semantic distance into one class. Finally, according to the hierarchical relationship of the clustering results, an ordered sequence from high correlation to low correlation is generated. In the example of high-density multilayer circuit boards, core similarity combinations of electrical performance are clustered into one class, those of physical structure into another class, and those of material properties into a third class. The sequence is arranged in the order of electrical performance (45%), physical structure (35%), and material properties (20%).

[0039] Step S1139: Based on the detection data type to which the core similarity combination belongs in the ordered core similarity combination sequence, establish logical associations between different types of core similarity combinations and generate a set of logical relationships between combinations.

[0040] This study analyzes the detection data types of core similar combinations in ordered sequences and establishes three types of logical associations: causal association (C), constrained association (R), and cooperative association (S). Causal association indicates that a change in one combination leads to a change in another, such as the association between on-resistance data and temperature drift coefficient characteristics (C type). Constrained association indicates that one combination provides a range limit for another, such as the association between aperture size data and graphic distortion rate characteristics (R type). Cooperative association indicates that two combinations jointly affect the same quality inspection index, such as the association between metallization layer thickness data and adhesion characteristics (S type). Logical associations are represented by a directed graph, where nodes are core similar combinations, directed edges represent the association type, and edge labels indicate the association strength (high / medium / low).

[0041] Step S11310: The ordered core similarity combination sequence is fused with the representation association relationship between different core similarity combinations to generate fused dimensional representation matching information, and the fused dimensional representation matching information is determined as the initial matching relationship of the dimensional representation content.

[0042] An ordered sequence and a logical association graph are fused using a graph fusion algorithm to generate a fused graph structure that includes node order and edge relationships. The fusion process employs an adjacency matrix representation, where rows represent core similar combinations in the ordered sequence, columns represent target combinations in the logical association, and matrix elements represent association type and strength. The fused dimensional representation of the matching information comprises three parts: a core matching relationship table (recording the correspondence between inherent dimensions and preset feature dimensions), an association strength matrix (recording the degree of influence between matching relationships), and a sequence priority list (recording the processing order of matching relationships). This initial matching relationship is stored as a JSON file, where the core matching relationship table contains 12 matching records, covering the main dimensions of the high-density multilayer circuit board detection data.

[0043] Step S114: Based on the initial matching relationship of the dimensional representation content, mark each inherent data dimension of the circuit board quality inspection test data set with the corresponding preset detection feature dimension of the deep learning quality inspection analysis model to generate dimension node marking information.

[0044] A dimensional node network is constructed using a graph database (Neo4j). Inherent data dimensions and preset detection feature dimensions are used as network nodes, and initial matching relationships are used as directed edges. Each node contains four attributes: a unique identifier (UUID), a dimension type label (inherent / preset), a dimension name, and a dimension identifier. Each edge contains four attributes: a source node ID, a target node ID, a matching strength (0-1), and a matching type (core / minor). In the high-density multilayer circuit board example, this network contains 28 inherent data dimension nodes and 15 preset feature dimension nodes, forming 32 directed edges (12 core matching edges and 20 minor matching edges). Node labeling is implemented using the graph database's CREATE statement, generating a unique node instance for each dimension and setting its attribute values.

[0045] Step S115: Bind the dimension node tagging information to the corresponding inherent data dimension and the corresponding preset detection feature dimension to generate dimension node binding information. Each dimension node binding information corresponds to a set of matched inherent data dimensions and preset detection feature dimensions.

[0046] For each core matching edge, dimension node binding information is generated, comprising three parts: a complete description of the inherent dimension (dimension identifier, attribute list, data example), a complete description of the preset feature dimension (dimension identifier, attribute list, feature extraction method), and matching relationship metadata (matching timestamp, matching algorithm version, matching confidence level). Taking the binding information of the excitation current level dimension (ER-CC-002) of conduction resistance and the test current intensity sub-content of the path resistance stability feature as an example, the inherent dimension description includes four attributes of "excitation current level," the preset feature description includes three feature parameters of "test current intensity," and the matching relationship metadata records a matching confidence level of 0.91. The binding information is stored in XML format, with the data structure defined through XSDschema to ensure format standardization.

[0047] Step S116: Classify all generated dimension node binding information according to the detection data type of the circuit board quality inspection test data set, and generate a subset of dimension node binding information divided by detection data type.

[0048] The binding information was categorized into three main types (electrical performance, physical structure, and material properties), with each type forming a subset. The electrical performance subset contained 12 binding information entries (43%), the physical structure subset contained 9 (32%), and the material properties subset contained 7 (25%). Within each subset, the data was sorted by the importance of each dimension. The importance was calculated using the Analytic Hierarchy Process (AHP). Five quality inspection experts were invited to conduct pairwise comparisons of the importance of each dimension, constructing a judgment matrix and calculating the weight vector. The classification results were stored in a file directory structure, with three subdirectories (EP, PS, MC) created under the main directory to store the corresponding type of binding information XML files.

[0049] Step S117: For each subset of dimension node binding information divided by the detection data type, establish the association relationship between the binding information of each dimension node within it, and generate a set of dimension association relationships corresponding to each subset of dimension node binding information divided by the detection data type.

[0050] Within the electrical performance type subset, the relationships between bound information are analyzed, establishing two types of relationships: data flow relationships (D) and parameter dependency relationships (P). Data flow relationships indicate that the output of one dimension serves as the input of another dimension, such as the on-resistance data flowing to the temperature drift coefficient feature (D type). Parameter dependency relationships indicate that a parameter in one dimension affects the calculation of another dimension, such as the excitation current range parameter affecting the resistance measurement accuracy feature (P type). Each relationship contains four attributes: source bound information ID, target bound information ID, relationship type, and influence coefficient (0-1). A total of 15 relationships are generated for the electrical performance subset, including 9 D-type and 6 P-type relationships, which are visualized through a directed graph.

[0051] Step S118: Merge the set of dimension association relationships corresponding to the subsets of dimension node binding information divided by the data type of detection, and generate the integrated set of dimension association relationships.

[0052] A set merging algorithm is employed to combine the dimensional relationships of the three subset types into a single set, removing duplicate records (determined by the combination of source ID and target ID) and supplementing cross-type relationships. Cross-type relationships primarily include: relationships between electrical properties and physical structures (e.g., the relationship between conduction resistance and hole wall roughness), relationships between physical structures and material properties (e.g., the relationship between dielectric layer thickness and dielectric constant), and relationships between electrical properties and material properties (e.g., the relationship between contact resistance and coating material). The merged set contains 42 relationships, forming a network covering all detection types. The merging process is implemented using Python's set data structure, leveraging the deduplication property of sets to ensure the uniqueness of the relationships.

[0053] Step S119: Merge the integrated set of dimensional relationships with all dimensional node binding information to generate fused dimensional mapping binding information, which includes all matching relationships and relationships.

[0054] Graph fusion technology integrates the binding information of dimensional nodes and the total set of dimensional relationships into a complete dimensional mapping graph. This graph comprises three layers: a node layer (inherent dimensional nodes and preset feature dimensional nodes), a binding layer (core matching edge layers), and an association layer (logical association edge layers). The node layer stores the basic attributes of the dimensions, the binding layer stores the matching relationships between inherent dimensions and feature dimensions, and the association layer stores the logical relationships between different dimensions. In the example of a high-density multilayer circuit board, this fused graph contains 43 nodes, 32 binding edges, and 42 association edges, with a total data size of approximately 2MB. The fused information is stored in GraphML format, which can be directly read and visualized by mainstream graph analysis tools (such as Gephi).

[0055] Step S1110: Determine the fused dimension mapping binding information as the binding association set between the data dimensions of the circuit board quality inspection test data set and the detection feature dimensions of the deep learning quality inspection analysis model.

[0056] The fused dimension mapping binding information undergoes integrity verification, checking for unmatched key dimensions (by comparing with a pre-defined list of key dimensions), closed loops in relationships (using a graph loop detection algorithm), and conformity of data format (using XMLSchema validation). Upon successful verification, this information is formally established as the binding association set, serving as the basis for subsequent model dimension expansion. During the verification process, it was found that the "dielectric constant temperature coefficient" dimension in the material properties class did not find a matching pre-defined feature dimension; it was marked as a dimension to be expanded and recorded in the dimension expansion requirement list. The final generated binding association set includes a GraphML format dimension mapping graph file and a verification report document. The verification report shows a key dimension matching rate of 92%, meeting the model input requirements.

[0057] Step S120: Based on the binding association set, drive the deep learning quality inspection analysis model to perform dynamic expansion processing of the detection feature dimensions, and obtain the expanded detection feature dimension set of the deep learning quality inspection analysis model.

[0058] The dimension expansion module of the deep learning quality inspection analysis model is launched, using the dimensions to be expanded marked in the bound association set as input. This module adopts a transfer learning framework, based on a pre-trained feature generation model (trained on a general industrial quality inspection dataset), and fine-tunes it for specific inspection dimensions in the circuit board field. In the example of high-density multilayer circuit boards, the dimensions to be expanded include 3 dimensions of material properties and 2 dimensions of physical structure, for a total of 5 dimensions to be expanded. The dimension expansion process is divided into three stages: requirements analysis, feature generation, and verification and optimization. Each stage generates corresponding intermediate results and log files to ensure traceability.

[0059] Step S121: Parse the dimension mapping binding information in the binding association set, extract the inherent data dimensions of the circuit board quality inspection test data set that do not match the preset detection feature dimensions of the deep learning quality inspection analysis model, and generate a set of unmatched inherent data dimensions.

[0060] The GraphML file containing the bound association set was queried using the SPARQL query language to filter out all intrinsic data dimension nodes without bound edges. The query condition was set to: the out-degree of the intrinsic dimension node is 0 (no matching preset feature dimension). In the example of a high-density multilayer circuit board, the query results returned 5 unmatched intrinsic dimensions: dielectric constant temperature coefficient (MP-DK-001), coefficient of thermal expansion (MP-CTE-002), plating hardness (MP-HD-003), blind via perpendicularity (PS-VT-004), and dielectric bubble rate (PS-BL-005). The identifiers, names, and attribute descriptions of the above dimensions were organized into a set of unmatched intrinsic data dimensions and stored as a CSV file containing 5 columns: dimension ID, dimension name, data type, unit, and importance score.

[0061] Step S122: For each inherent data dimension in the set of unmatched inherent data dimensions, extract all attribute information of its associated detection data to generate a set of refined attribute information for each unmatched inherent data dimension.

[0062] Step S1221: Select one inherent data dimension from the set of unmatched inherent data dimensions as the dimension to be processed, extract the original characterization information of the circuit board quality inspection test data corresponding to the dimension to be processed, and obtain the description content of all test data under the dimension to be processed.

[0063] Using the temperature coefficient of dielectric constant (MP-DK-001) as the dimension to be processed, detection data records for this dimension were extracted from the raw database of the MES system. The raw characterization information includes four types of data: test report PDF documents, raw sampling data CSV files, equipment operation log TXT files, and image detection result PNG files. The CSV file contains 1000 sampling records, each containing four fields: test temperature, applied frequency, dielectric constant value, and loss factor. The PDF report contains three sections: test method description, equipment calibration certificate, and uncertainty analysis. Data integration tools were used to convert the raw information in different formats into a unified text description, forming the raw characterization information database for this dimension.

[0064] Step S1222: The original representation information of the dimension to be processed is split according to the acquisition method of the detection data to generate multiple sets of original representation sub-information with different acquisition methods. Each set of original representation sub-information corresponds to the representation content under a data acquisition method.

[0065] The temperature coefficient of dielectric constant was measured using three acquisition methods: capacitor bridge method (CBM), resonant cavity method (RCM), and time-domain reflectometry (TDR). The raw characterization information was split according to the acquisition method, resulting in three sub-information sets: CBM sub-information (containing 200 data points), RCM sub-information (containing 300 data points), and TDR sub-information (containing 500 data points). Each sub-information set includes four parts: equipment parameters for that acquisition method (e.g., CBM bridge type, RCM cavity length), environmental conditions (temperature control accuracy, humidity range), data processing algorithm (e.g., FFT transform parameters), and original data samples. The splitting process was implemented through file directory classification. Three subdirectories were created under the directory of the dimension to be processed to store the sub-information files for different acquisition methods.

[0066] Step S1223: For each set of original characterization sub-information with different acquisition methods, by parsing its data format and content, identify the parameter items and their values ​​recorded in it that are related to circuit board quality inspection, and generate the parameter item set of the corresponding original characterization sub-information.

[0067] Taking CBM sub-information as an example, a combination of regular expression matching and natural language processing was used to parse the parameter items. First, the PDF test report was extracted using the PyPDF2 library. Then, the BERT model was used to perform Named Entity Recognition (NER) on the text to identify the parameter item names (such as "test frequency" and "electrode area"). Finally, keyword matching was used to extract the value range of the corresponding parameters from the CSV data (e.g., the test frequency range is 1kHz to 1GHz). A total of eight parameter items were identified in the CBM sub-information: test frequency, electrode type, applied voltage, temperature step, equilibration time, number of measurement points, data averaging frequency, and calibration cycle. Each parameter item records five attributes: name, data type, unit, typical value, and value range, stored as a JSON-formatted parameter item collection.

[0068] Step S1224: Classify the parameter item set of each group of original representation sub-information according to the type of parameter item, and generate a sub-category parameter item subset. Each sub-category parameter item subset corresponds to a class of parameter items of the same type.

[0069] The eight parameters of the CBM sub-information are categorized into four subsets: electrical parameters (test frequency, applied voltage), physical parameters (electrode type, electrode area), timing parameters (temperature step, equilibration time, number of measurement points), and system parameters (data averaging times, calibration cycle). The electrical parameters subset contains parameters related to electrical signals, the physical parameters subset contains parameters related to test fixtures, the timing parameters subset contains parameters related to the test process, and the system parameters subset contains parameters related to equipment status. The classification is based on the physical meaning of the parameters and their role in the testing process. Automatic classification is achieved through expert-defined rules, which are stored in the form of a decision tree and can be edited and updated using visualization tools.

[0070] Step S1225: For each subset of parameter items of a subtype, generate its corresponding metadata description. The metadata description includes the name, data type, value range and physical unit of the parameter item, and generate metadata description information.

[0071] Metadata descriptions are generated for a subset of electrical parameters. The metadata for the "Test Frequency" parameter includes: name (test frequency), data type (continuous numerical), value range (1kHz-1GHz), physical unit (Hertz), accuracy class (±0.1%), and resolution (1Hz). The metadata for the "Applied Voltage" parameter includes: name (applied voltage), data type (discrete numerical), value range (1V-10V), physical unit (volts), accuracy class (±0.05V), and resolution (0.01V). The metadata descriptions are extended from the DublinCore metadata standard, adding accuracy class and resolution attributes specific to the industrial testing field. The metadata description information is stored in RDF format for easy parsing and utilization by semantic web applications.

[0072] Step S1226: Summarize the metadata description information corresponding to the original representation sub-information under all acquisition methods to generate a preliminary refined attribute information set of the dimension to be processed. The preliminary refined attribute information set of the dimension to be processed covers the parameter items of all acquisition methods.

[0073] Metadata descriptions from CBM, RCM, and TDR acquisition methods are aggregated to obtain a preliminary set of refined attribute information for the dielectric constant temperature coefficient dimension. This set contains 23 parameter items (8 from CBM, 7 from RCM, and 8 from TDR), with duplicate parameter items (such as test temperature) retaining the description with the widest value range. Each parameter item includes three parts: acquisition method source, metadata description, and importance score (1-5 points). The importance score is assessed by domain experts based on the parameter's impact on the quality inspection results; for example, "test temperature" has an importance score of 5 points (highest), and "data average frequency" has a score of 2 points. The aggregation process is implemented using a UNION operation in a relational database, removing duplicate records and sorting them by importance score.

[0074] Step S1227: Extract the industry standard characterization rules for circuit board quality inspection and testing, and adjust the terminology and form of the preliminary refined attribute information set of the dimension to be processed according to the industry standard characterization rules for circuit board quality inspection and testing to generate an aligned refined attribute information set.

[0075] Referring to relevant clauses in IPC-A-600H (Printed Circuit Board Appearance Inspection Standard) and IPC-TM-650 (Printed Circuit Board Test Methods Manual), the preliminary refined attribute information set was standardized and adjusted. The adjustments included: standardization of parameter item names (e.g., unifying "applied voltage" as "excitation voltage"), standardization of data types (e.g., unifying "discrete numerical type" as "enumerated type"), standardization of units (e.g., unifying "Hertz" as "Hz"), and formatting of value ranges (e.g., representing "1kHz-1GHz" as "[1e3,1e9]Hz"). For example, the "pulse width" parameter item in the TDR acquisition method was adjusted to "rise edge time" according to the IPC standard, the data type was changed from "string type" to "time type", and the unit was changed from "ns" to "μs". The adjusted refined attribute information set contains 20 standardized parameter items, reducing 3 non-standard parameter items.

[0076] Step S1228: Remove completely duplicate entries and entries that are unrelated to the preset list of key dimensional attributes from the aligned set of refined attribute information to generate a simplified set of refined attribute information.

[0077] The pre-defined list of key attributes for each dimension was defined by PCB quality control experts and includes 12 core attributes that significantly impact feature extraction: temperature coefficient, frequency characteristics, humidity sensitivity, test accuracy, repeatability, stability, environmental dependence, sample size requirements, equipment compatibility, data resolution, calibration cycle, and security level. The aligned 20 parameter items were compared with this list, retaining matching items and removing completely duplicate entries (e.g., only one "test temperature" parameter item was retained for different acquisition methods). This resulted in 12 key parameter items, such as dielectric constant temperature coefficient (25℃-125℃), frequency characteristics (1kHz-10GHz), and humidity sensitivity (RH 30%-90%). The simplification process was implemented using Python's Pandas library, employing the `drop_duplicates` and `isin` methods for deduplication and filtering.

[0078] Step S1229: Arrange the simplified set of refined attribute information according to the importance of the parameter items to generate an ordered set of refined attribute information, in which the representation content corresponding to the importance of the parameter items is arranged in order.

[0079] The objective weight of each parameter is calculated using the entropy weight method, and its importance is determined based on the degree of variation of the parameter in historical detection data. The greater the degree of variation (the smaller the entropy value), the higher the weight. In the example of the dielectric constant temperature coefficient dimension, "temperature coefficient (25℃-125℃)" has the smallest entropy value (0.12) and the highest weight (0.23); "calibration cycle" has the largest entropy value (0.87) and the lowest weight (0.05). The 12 parameter items are arranged in descending order of weight to generate an ordered set of refined attribute information. The arrangement result is stored as a JSON array, with each element containing three fields: parameter name, weight value, and metadata description, which facilitates subsequent processing by the feature generation model according to importance.

[0080] Step S12210: Determine the ordered set of refined attribute information as the set of refined attribute information for the unmatched inherent data dimension, process all dimensions in the set of unmatched inherent data dimensions in sequence, and generate a set of refined attribute information for all unmatched inherent data dimensions.

[0081] For the five dimensions (temperature coefficient of dielectric constant, coefficient of thermal expansion, coating hardness, blind via perpendicularity, and dielectric bubble rate) in the unmatched inherent data dimension set, steps S1221 to S1229 were executed one by one to generate their respective ordered and refined attribute information sets. The processing result for each dimension includes a JSON-formatted attribute set file and a processing log file. The log file records the processing start time, processing duration, called utility functions, expert adjustment records, and other information. In the processing of the coefficient of thermal expansion dimension, it was found that there were no clear characterization rules in the industry standard; therefore, internal company standards were used for adjustment, and the reasons and basis were noted in the log. The final generated refined attribute information set for all unmatched dimensions is approximately 1.2MB in size, containing five JSON files and five log files.

[0082] Step S123: Input the set of refined attribute information for each unmatched inherent data dimension into the dimension generation module of the deep learning quality inspection analysis model, drive the dimension generation module of the deep learning quality inspection analysis model to construct new detection feature dimensions, and generate new detection feature dimensions corresponding to the unmatched inherent data dimensions.

[0083] The dimension generation module of the deep learning quality inspection analysis model adopts a variational autoencoder (VAE) architecture, which includes an encoder (3 fully connected layers), a decoder (3 fully connected layers), and a feature mapping layer (1 convolutional layer). The encoder compresses the input set of refined attribute information (a metadata vector with 12 parameter items) into a 32-dimensional latent feature vector. The decoder reconstructs the latent feature vector into attribute information and optimizes the model parameters using a loss function (reconstruction loss + KL divergence). The feature mapping layer maps the latent feature vector into feature dimension descriptions that meet the model input requirements. During the training phase, a dataset containing 100,000 sets of industrial material property data is used for pre-training, followed by fine-tuning with 2,000 sets of labeled data from the circuit board domain. The optimizer used is Adam, with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs. During the inference phase, the input set of refined attribute information (dielectric constant and temperature coefficient) is used. The encoder outputs a 32-dimensional vector, which, after processing by the feature mapping layer, generates descriptions of new detection feature dimensions, including feature name, dimension size, data type, and extraction method.

[0084] Step S124: Extract the representation rules of the preset detection feature dimension of the deep learning quality inspection analysis model, adjust the representation form of the new detection feature dimension according to the representation rules of the preset detection feature dimension of the deep learning quality inspection analysis model, and generate a new detection feature dimension that is consistent with the preset detection feature dimension representation rules.

[0085] The predefined feature dimension representation rules are stored in the model's configuration file, including three categories of rules: naming conventions, data structure conventions, and interface conventions. The naming convention requires feature dimension names to follow the format "feature family-core attribute-serial number," such as "MCF-DK-001" (Material Compatibility Feature Family-Dielectric Constant-001). The data structure convention requires feature dimension descriptions to include six fields: name, ID, type, shape, dimension, and precision. The interface convention defines the input and output format of feature data as ProtocolBuffers. New detection feature dimensions output by the dimension generation module are adjusted according to the above rules: the name is changed from "Dielectric Constant Temperature Characteristics" to "MCF-DK-001," the data type is changed from "Custom" to "Floating-Point Array," and the shape is changed from "

[100] " to "[1,100]" to meet the batch dimension requirements of the model input. The adjusted new detection feature dimension description is verified through JSONSchema to ensure complete consistency with the predefined dimension representation.

[0086] Step S125: Generate dimension identifiers for all new detection feature dimensions after unification of form, and add a unique dimension identifier for each new detection feature dimension. The unique dimension identifier of the new detection feature dimension is distinct from the dimension identifier of the preset detection feature dimension and has no duplication.

[0087] Dimension identifiers are generated using the UUIDv4 algorithm to ensure global uniqueness. Simultaneously, for ease of manual identification, a readability identifier is added to the UUID, formatted as "NEW-Feature Family-Serial Number," such as "NEW-MCF-001" (New Material Compatibility Feature-001). The unique dimension identifier is stored in two parts: a machine-readable UUID (for internal model references) and a human-readable feature code (for logs and reports). In the high-density multilayer circuit board example, the UUIDs for the five new detection feature dimensions are 32-bit hexadecimal strings, with readability identifiers sequentially from NEW-MCF-001 to NEW-MCF-003 (material property class) and NEW-SIF-004 to NEW-SIF-005 (physical structure class). After generation, the dimension identifiers are stored in association with the feature dimension description information to ensure a one-to-one correspondence.

[0088] Step S126: Extract the detection process sequence of the circuit board quality inspection test, and arrange the new detection feature dimension with exclusive dimension identifier and the preset detection feature dimension of the deep learning quality inspection analysis model according to the detection process sequence of the circuit board quality inspection test, so as to generate an ordered detection feature dimension sequence.

[0089] The standard procedure for PCB quality inspection is as follows: Incoming material inspection → Inner layer fabrication → Lamination → Drilling → Electroplating → Outer layer fabrication → Surface treatment → Finished product inspection. Each inspection stage corresponds to a specific inspection feature dimension, such as the conductivity reliability feature family for the inner layer fabrication stage and the material compatibility feature family for the lamination stage. Following this procedure, preset and new inspection feature dimensions are arranged in a mixed manner: Incoming material inspection stage (2 preset dimensions) → Inner layer fabrication stage (3 preset dimensions) → Lamination stage (2 preset dimensions + 3 new dimensions) → Drilling stage (2 preset dimensions + 2 new dimensions) → Subsequent stages (6 preset dimensions). The resulting ordered sequence of inspection feature dimensions contains 20 dimensions (15 preset + 5 newly generated), arranged from left to right according to the chronological order of the inspection process. Each dimension is labeled with its corresponding inspection stage and importance score.

[0090] Step S127: According to the quality inspection analysis logic, establish a correlation between the new detection feature dimension and the preset detection feature dimension in the ordered detection feature dimension sequence, and generate a set of inter-dimensional correlation relationships.

[0091] The quality control analysis logic is represented by expert-defined production rules, such as "if the dielectric constant temperature coefficient exceeds the threshold, the coefficient of thermal expansion needs to be checked accordingly." Based on these rules, relationships are established between dimensions in the ordered sequence, including three types: causal relationships (A causes B), constraint relationships (A restricts the value of B), and synergistic relationships (A and B jointly affect C). During the lamination stage, the newly generated dielectric constant temperature coefficient dimension (NEW-MCF-001) establishes a causal relationship with the preset interlayer peel strength characteristic dimension (SIF-PE-002), with a relationship strength score of 0.85; and a synergistic relationship with the preset thermal stress distribution characteristic dimension (MCF-TS-003), with a relationship strength score of 0.78. Each relationship contains five fields: source dimension ID, target dimension ID, relationship type, strength score, and rule basis, and is stored as a CSV format file containing a collection of inter-dimensional relationships.

[0092] Step S128: Integrate the dimensional correlation between the new detection feature dimension and the preset detection feature dimension into the ordered detection feature dimension sequence to generate a detection feature dimension sequence with dimensional correlation. The detection feature dimension sequence with dimensional correlation includes the representation content, identifiers and interrelationships of all dimensions.

[0093] A Directed Acyclic Graph (DAG) data structure is used to represent the sequence of detection feature dimensions with relationships. Nodes represent dimensions (preset or newly generated), and directed edges represent relationships. In the DAG, node attributes include dimension identifier, name, type (preset / new), and stage; edge attributes include relationship type, strength, and rule ID. A topological sorting algorithm ensures the acyclicity of the DAG. If a cycle is detected (e.g., A→B→C→A), the strength of the weakest relationship edge is reduced until the cycle disappears. In the example of a high-density multilayer circuit board, this DAG contains 20 nodes and 35 directed edges, of which 12 edges involve new detection feature dimensions (accounting for 34%). The sequence with relationships is stored as a Graphviz DOT format file, which can generate a visualized directed graph.

[0094] Step S129: Based on the set of inter-dimensional relationships, perform topological sorting on the detection feature dimension sequence with dimensional relationships to generate a topologically ordered detection feature dimension sequence as the complete detection feature dimension arrangement result.

[0095] The Kahn algorithm is used to perform topological sorting on the sequence of detection feature dimensions with correlations. The algorithm steps are as follows: 1) Calculate the in-degree of all nodes; 2) Add nodes with an in-degree of 0 to the queue; 3) Remove nodes from the queue and add them to the sorting result; 4) Delete the node and its outgoing edges, and update the in-degree of adjacent nodes; 5) Repeat steps 2-4 until the queue is empty. In the example of high-density multilayer circuit boards, the topological sorting result is basically consistent with the detection process order, only adjusting the order of two strongly correlated dimensions (adjusting the thermal expansion coefficient dimension after the dielectric constant temperature coefficient dimension). The sorting result generates two outputs: a topologically ordered list of dimension IDs (JSON array) and a sorting process log (recording the in-degree changes and node processing order). Topological sorting ensures that when processing feature dimensions, the model processes the preceding dimensions first, and then processes the subsequent dimensions that depend on it.

[0096] Step S1210: Determine the complete arrangement of detection feature dimensions as the expanded set of detection feature dimensions of the deep learning quality inspection analysis model.

[0097] The topologically ordered sequence of detection feature dimensions undergoes final validation, checking for dimensional completeness (whether it includes all preset and newly generated dimensions), consistency of associations (whether it matches the set of associations between dimensions), and rationality of the topological order (whether it conforms to the detection process logic). Once validation is successful, this sequence is designated as the expanded set of detection feature dimensions, replacing the model's original preset set of feature dimensions. In the high-density multilayer circuit board example, the expanded set contains 20 feature dimensions (15 preset + 5 newly generated), an increase of 5 dimensions compared to the original set, covering all unmatched inherent data dimensions. The final expanded set of detection feature dimensions includes three files: a topologically ordered list of dimensions (JSON), a dimension association graph (DOT), and a dimension description document (PDF), serving as the basis for subsequent feature extraction by the model.

[0098] Step S130: Use the expanded detection feature dimension set to perform deep feature mining processing on the circuit board quality inspection test data set to obtain the deep mining feature set of the circuit board quality inspection test data set.

[0099] The expanded set of detection feature dimensions is loaded into the feature mining module of the deep learning quality inspection analysis model. This module adopts a hybrid deep learning architecture, including a Convolutional Neural Network (CNN) submodule (for spatial feature extraction), a Long Short-Term Memory (LSTM) submodule (for temporal feature extraction), and a Graph Convolutional Network (GCN) submodule (for relational feature extraction). In the example of high-density multilayer circuit boards, the circuit board quality inspection test dataset contains the detection data of 1000 circuit boards, each with 28 detection indicators, and the total data volume is approximately 5GB. The deep feature mining process is divided into four stages: data preprocessing, multi-channel feature extraction, cross-channel feature fusion, and feature selection. Each stage has clear inputs, outputs, and evaluation indicators.

[0100] Step S131: Input the expanded detection feature dimension set as the mining basis into the feature mining module of the deep learning quality inspection analysis model, and set the mining dimension standard for the feature mining module of the deep learning quality inspection analysis model that matches the circuit board quality inspection test data.

[0101] The feature mining module reads the JSON file containing the expanded set of detection feature dimensions, parses the metadata information (name, ID, type, shape, and dimension) of the 20 feature dimensions, and converts it into feature descriptors for the model. Each feature descriptor contains four parameters: dimension size (e.g., a 100-dimensional vector), data type (e.g., float32), preprocessing method (e.g., normalization, standardization), and extraction algorithm identifier (e.g., CNN, LSTM). For example, the feature descriptor for the dielectric constant temperature coefficient dimension (NEW-MCF-001) is set to: dimension size 100, data type float32, preprocessing method Z-Score normalization, and extraction algorithm GCN. The mining dimension standards are stored as a configuration file (YAML format) for the model, specifying the processing requirements for each feature dimension to ensure consistency between the feature mining process and the dimension standards.

[0102] Step S132: Split the circuit board quality inspection test data set according to the type of detection data to generate multiple circuit board quality inspection test data subsets. Each circuit board quality inspection test data subset corresponds to a type of detection feature dimension in the expanded detection feature dimension set.

[0103] Based on the classification of test data types (electrical performance, physical structure, and material properties) and the expanded set of test feature dimensions, the original circuit board quality inspection test data set was split into three subsets: an electrical performance data subset (corresponding to 7 feature dimensions), a physical structure data subset (corresponding to 8 feature dimensions), and a material property data subset (corresponding to 5 feature dimensions). Each subset contains test data for all circuit boards of that type, and the data format was converted to a tensor format (N×C×H×W) required by the model. In the example of a high-density multilayer circuit board, the shape of the electrical performance data subset is (1000×3×256×256), the physical structure data subset is (1000×4×512×512), and the material property data subset is (1000×2×128×128). The splitting process was implemented using Python's NumPy library, using array slicing and dimension rearrangement operations to convert the original data into tensors that meet the model input requirements.

[0104] Step S133: Adapt the data and dimensions of each circuit board quality inspection test data subset to the corresponding detection feature dimension, and import the circuit board quality inspection test data subset into the mining channel of the corresponding detection feature dimension for feature mining.

[0105] The feature mining module comprises three parallel mining channels, corresponding to subsets of electrical performance, physical structure, and material properties data, respectively. Each channel includes a data adaptation layer and a feature extraction network: the data adaptation layer transforms the input data subset into the format required by the feature extraction network (e.g., adjusting the number of channels, normalizing the data range); the feature extraction network selects the appropriate network architecture based on the type of feature dimension (CNN for image-based data, LSTM for time-series data, and GCN for correlation-based data). In the high-density multilayer circuit board example, the electrical performance data subset (time-series) is imported into the LSTM mining channel, the physical structure data subset (image-based) is imported into the CNN mining channel, and the material properties data subset (correlation-based) is imported into the GCN mining channel. The data import process is implemented through the model's input interface, with each channel having an independent input queue and preprocessing thread to ensure efficient data loading.

[0106] Step S134: In each detection feature dimension mining channel, the imported circuit board quality inspection test data subset is processed layer by layer through multiple cascaded feature extraction layers of the deep learning quality inspection analysis model. The output of each layer is used as the input of the next layer to generate a multi-layer feature output set of the corresponding circuit board quality inspection test data subset.

[0107] Step S1341: Take the subset of circuit board quality inspection test data imported into the detection feature dimension mining channel as the initial mining data, extract the surface data representation content of the initial mining data, and generate the surface feature set of the initial mining data. The surface feature set of the initial mining data reflects the intuitive representation attributes of the data.

[0108] Taking the import of a subset of physical structure data (images) into a CNN for channel mining as an example, the initial mining data consists of 1000 optical inspection images of circuit boards with a resolution of 512×512. Surface feature extraction is achieved through the first two layers of the CNN: the first layer (convolutional layer) uses 32 convolutional kernels of 3×3, with a stride of 1, padding of 1, and the ReLU activation function, outputting a 32×512×512 feature map; the second layer (pooling layer) uses a 2×2 max-pooling kernel with a stride of 2, outputting a 32×256×256 feature map. The surface feature set contains the output feature map of this pooling layer, reflecting the intuitive representation attributes of the image such as edges and textures. Each feature map corresponds to a specific visual mode of the image (such as blind hole edge mode, line corner mode), and the feature value represents the response intensity of that mode. The surface feature set is stored as a 4D tensor (number of samples × number of channels × height × width), with a shape of (1000×32×256×256) in the example.

[0109] Step S1342: Input the surface feature set of the initial mining data into the first feature extraction layer of the mining channel, process the surface features through the first feature extraction layer, mine the shallow hidden features behind the surface features, and generate the shallow hidden feature set of the initial mining data.

[0110] The first feature extraction layer in the CNN's channel mining process is a residual block (ResBlock), containing two convolutional layers and one skip connection: First convolutional layer (64 3×3 kernels) → batch normalization → ReLU → Second convolutional layer (64 3×3 kernels) → batch normalization → skip connection (adding the input features to the output of the second convolutional layer) → ReLU. The input surface feature set (32×256×256) is processed by this residual block, outputting a 64×256×256 shallow latent feature set. These shallow latent features mine local structural features of the image, such as the circularity of blind holes and the continuity of lines, which possess stronger semantic information than surface features. The output of this layer is used to generate a heatmap using feature visualization tools, displaying the image regions of interest to the model, facilitating manual review of the reasonableness of feature extraction.

[0111] Step S1343: The shallow hidden feature set of the initial mining data is transferred from the first feature extraction layer of the mining channel to the second feature extraction layer. The shallow hidden features are processed in depth by the second feature extraction layer to mine the middle hidden features behind the shallow hidden features, and the middle hidden feature set of the initial mining data is generated.

[0112] The second feature extraction layer consists of two cascaded residual blocks, each with 128 convolutional kernels. The first residual block transforms the shallow latent features (64×256×256) into a 128×256×256 feature map, while the second residual block maintains the same number of channels and outputs a 128×256×256 mid-level latent feature set. The mid-level latent features extract component-level features from the image, such as the arrangement of blind aperture arrays and the inter-layer alignment of lines, exhibiting stronger structural characteristics. This layer applies dropout technology (dropout rate 0.3) to prevent overfitting, randomly discarding the outputs of some neurons during the training phase. The mid-level latent feature set has a shape of (1000×128×256×256), with the same resolution as the shallow latent features but double the number of channels, resulting in enhanced feature representation.

[0113] Step S1344: The middle-level hidden feature set of the initial mining data is transferred from the second feature extraction layer of the mining channel to the third feature extraction layer. The middle-level hidden features are further processed by the third feature extraction layer to mine the deep hidden features behind the middle-level hidden features and generate the deep hidden feature set of the initial mining data.

[0114] The third feature extraction layer consists of a residual block (256 convolutional kernels) and a global average pooling layer. The residual block transforms the mid-layer latent features (128×256×256) into a 256×256×256 feature map. The global average pooling layer performs average pooling on the feature map of each channel, outputting a 256×1×1 feature vector (flattened to a 256-dimensional vector). The deep latent features are abstract semantic features, such as "blind via quality level" and "circuit defect risk," which are directly related to the quality inspection conclusions. In the example of a high-density multilayer circuit board, the shape of the deep latent feature set is (1000×256), with each sample represented as a 256-dimensional feature vector. The output of this layer serves as the input for subsequent feature fusion and is also saved as an NPY format file for model interpretation and feature analysis.

[0115] Step S1345: The deep hidden feature set of the initial mining data is transferred from the third feature extraction layer of the mining channel to the subsequent feature extraction layer, and feature mining is performed in sequence according to the aforementioned feature processing method, gradually delving into the underlying representation of the data.

[0116] Depending on the complexity of the feature dimensions, subsequent feature extraction layers can contain 1-3 additional fully connected layers, with the number of neurons in each layer halved sequentially (256→128→64). Each fully connected layer is followed by batch normalization and a ReLU activation function, and the last layer outputs a 64-dimensional latent feature vector. The latent features represent the essential attributes of the data, such as the microstructural characteristics of the material and the fluctuation patterns of process parameters. These features cannot be defined manually and can only be automatically mined through deep learning. In the example of a high-density multilayer circuit board, the subset of physical structure data is processed through 5 layers of feature extraction (2 surface layers + 3 deep layers) to generate a 64-dimensional latent feature vector, with a feature set shape of (1000×64). The output features of each layer are saved to a temporary file for analyzing the abstraction process of features as the number of layers increases.

[0117] Step S1346: In each subsequent feature extraction layer, perform at least one of the following operations: convolution, pooling, or fully connected operations on the received feature set from the previous layer to generate the output feature set of the corresponding feature extraction layer.

[0118] Taking LSTM channel mining (processing electrical performance time-series data) as an example, the subsequent feature extraction layers consist of three LSTM layers and two fully connected layers: First LSTM layer (128 hidden units, return sequence = True) → Second LSTM layer (64 hidden units, return sequence = True) → Third LSTM layer (32 hidden units, return sequence = False) → Fully connected layer (64 neurons) → Fully connected layer (32 neurons). The operations performed in each layer are as follows: the LSTM layer recursively processes the time-series features to capture long-term dependencies; the fully connected layer performs a non-linear transformation on the hidden states output by the LSTM to extract higher-order features. The shape of the output feature set gradually transforms from the initial (1000×24×64) (24 time steps, 64 features) to the final (1000×32) feature vector. The parameters of each layer operation (such as the number of hidden units in the LSTM and the number of neurons in the fully connected layer) are pre-set according to the complexity of the feature dimensions and optimized using a validation set.

[0119] Step S1347: Record the mapping relationship between the feature set generated by each feature extraction layer and its input feature set, and generate inter-layer feature mapping relationship information. The inter-layer feature mapping relationship information reflects the transmission and evolution relationship between features of different extraction layers.

[0120] By analyzing the feature importance during gradient backpropagation, the mapping relationship between input and output features at each layer is recorded. For CNN layers, it records which input feature maps are combined by which convolutional kernels to form each output feature map; for LSTM layers, it records the dependency between the output hidden state and each time step in the input sequence; for fully connected layers, it records the weight connection strength between output neurons and input neurons. The inter-layer feature mapping information is stored as a sparse matrix (only recording non-zero weights) and a visual heatmap is generated to show the contribution of input features to output features. In the example of a high-density multilayer circuit board, the mapping relationship of the first convolutional layer in the CNN channel mining layer shows that the output feature map is mainly composed of edge and texture features of the input image, consistent with the expected visual feature extraction logic.

[0121] Step S1348: Summarize the feature sets generated by all feature extraction layers, including the surface feature set, the hidden feature sets of each layer, and the feature mapping relationship information between layers, to generate the feature extraction summary information of the initial mining data.

[0122] The output features of all feature extraction layers (surface features, shallow hidden features, mid-layer hidden features, deep hidden features, and bottom-layer hidden features) and inter-layer mapping relationships are summarized to form feature extraction summary information. This information includes three parts: a feature set list (recording the shape, data type, and storage path of each layer's features), a mapping relationship matrix (recording the weights and connections of inter-layer feature mappings), and feature statistics (recording the mean, variance, and sparsity of each layer's features). In the example of a high-density multilayer circuit board, the feature extraction summary information of the physical structure data subset includes 5 feature sets (approximately 8GB of data) and 4 inter-layer mapping relationship matrices (approximately 2GB of data). The summarization process is executed automatically by a script, generating a structured directory tree for easy and quick access to features at each layer later.

[0123] Step S1349: Arrange the feature extraction summary information of the initial mining data according to the hierarchical order of feature extraction to generate a multi-layer feature output set, which reflects the progressive extraction process of features from the surface to the bottom layer.

[0124] Following the hierarchical order of feature extraction (surface layer → shallow layer → middle layer → deep layer → bottom layer), the feature set list is sorted to generate a multi-layer feature output set. Each layer's features contain four metadata items: feature data file path, feature dimension, extraction timestamp, and confidence score. In the high-density multilayer circuit board example, the multi-layer feature output set of the physical structure data subset contains features at five levels, numbered from 1 (surface layer) to 5 (bottom layer), with the feature dimension gradually decreasing from 32×256×256 to 64. The multi-layer feature output set is stored as a JSON-formatted index file, recording the metadata and access paths of each layer's features. The model can load features from specific layers for subsequent processing as needed.

[0125] Step S135: Based on the preset feature importance threshold, filter the features in the multi-layer feature output set of each circuit board quality inspection test data subset, retain the features with importance higher than the feature importance threshold, and generate the core-level feature set of the corresponding circuit board quality inspection test data subset.

[0126] Feature importance is calculated using SHAP (SHapley Additive ex Planations) values, measuring the contribution of each feature to the model output. The preset feature importance threshold is set based on the number of feature dimensions and model complexity; in the high-density multilayer circuit board example, the threshold is set to the top 60% of all feature SHAP values. The multilayer feature output set of the physical structure data subset is filtered: 19 feature maps from the top 60% of surface features (32), 77 from shallow features (128), 77 from mid-level features (128), 154 from deep features (256), and 38 from bottom-level features (64). The filtered core-level feature set contains the retained features and their SHAP values ​​for each level, stored as a feature index list (recording the retained feature numbers) and a feature data file (containing only the retained feature data), reducing the amount of data for subsequent processing.

[0127] Step S136: Arrange the core-level feature set of all circuit board quality inspection test data subsets according to the dimensional order of the expanded detection feature dimension set to generate an ordered core-level feature sequence.

[0128] The expanded feature dimension set consists of 20 topologically ordered feature dimensions, with each data subset's core-level feature set corresponding to a portion of these dimensions. The core-level feature sets of the three subsets—electrical performance, physical structure, and material properties—are interleaved according to their dimensional order: dimensions 1-7 (electrical performance) → dimensions 8-15 (physical structure) → dimensions 16-20 (material properties). The resulting ordered core-level feature sequence contains core features across 20 feature dimensions, with each dimension's features being the core features selected from the multi-layer feature output set for that dimension. In the high-density multilayer circuit board example, the sequence length is 20 (number of dimensions) × 64 (average number of features per layer) = 1280 dimensions, and each sample is represented as a 1280-dimensional feature vector. The arrangement process is implemented using NumPy's `concatenate` function, concatenating the feature vectors of the three subsets according to their dimensional order.

[0129] Step S137: Input the ordered core hierarchical feature sequence into the feature fusion layer, and use the feature fusion layer to perform weighted combination and nonlinear transformation on the core hierarchical features from different detection feature dimensions to generate a fused hierarchical feature set.

[0130] The feature fusion layer employs an architecture combining attention mechanisms and autoencoders. First, a multi-head self-attention module (8 heads, each with 16 dimensions) calculates attention weights for different feature dimensions, reflecting the importance of that dimension to the quality inspection result. Then, an autoencoder (encoder: 1280→640→320; decoder: 320→640→1280) performs a non-linear transformation on the weighted feature sequence to extract cross-dimensional correlation features. Finally, a skip connection adds the input features to the decoder output, preserving the original feature information. The fused hierarchical feature set has the same shape as the input (1000×1280), but the correlation information between features is enhanced, and redundant information is suppressed. The output of this layer is displayed after dimensionality reduction (t-SNE) using a visualization tool. Samples from different quality inspection results form obvious clusters in the feature space, verifying the fusion effect.

[0131] Step S138: Input the fused hierarchical feature set into the attention mechanism layer, calculate the weight of each feature through the attention mechanism layer and weight the features to generate a weighted fused hierarchical feature set.

[0132] The attention mechanism layer comprises two sub-modules: spatial attention and channel attention. The spatial attention module (7×7 convolution) focuses on important regions in the feature map, while the channel attention module (global average pooling + fully connected layers) focuses on important feature channels. The output weights of the two modules are multiplied and then added to the input features to achieve adaptive feature weighting. In the example of a high-density multilayer circuit board, the spatial attention weights highlight the pads and blind vias of the circuit board, while the channel attention weights highlight feature channels related to conductivity and insulation. The shape of the weighted fusion layer feature set remains (1000×1280), but the distribution of feature values ​​is more concentrated on important features, significantly improving the signal-to-noise ratio. The attention weights of this layer are saved as image files to explain the regions and channels that the model focuses on.

[0133] Step S139: The weighted fusion-level feature set is reclassified according to the detection location of the circuit board quality inspection test, generating a fusion-level feature subset divided by the detection location. Each fusion-level feature subset divided by the detection location corresponds to a specific detection location of the circuit board.

[0134] The circuit board inspection area is divided into five regions: pad area, circuit area, blind via area, dielectric layer area, and edge area. Each region corresponds to specific inspection indicators and defect types. For example, the pad area focuses on pad diameter and adhesion, while the blind via area focuses on via diameter and perpendicularity. Weighted fused features are assigned to corresponding regions using a region mask matrix (a binary matrix associated with feature dimensions): elements with a value of 1 in the mask matrix indicate that the feature belongs to the corresponding region, and elements with a value of 0 indicate that it does not belong. In the example of a high-density multilayer circuit board, five fused hierarchical feature subsets are generated, each with a feature dimension of approximately 256 (1280 total dimensions / 5 regions) and a shape of (1000×256). The region division process is implemented through matrix multiplication, multiplying the feature matrix with the mask matrix to extract region-related features.

[0135] Step S1310: Summarize and integrate all the fusion-level feature subsets divided by the detection parts to generate a feature set covering all detection parts and all detection dimensions of the circuit board. The feature set covering all detection parts and all detection dimensions of the circuit board is determined as the deep mining feature set of the circuit board quality inspection test data set.

[0136] Five fusion-level feature subsets, divided by location, are aggregated and then concatenated (along the feature dimensions) to generate a complete deep-mining feature set. The concatenation order follows the importance of the detected location (pad area → circuit area → blind via area → dielectric layer area → edge area), ensuring that features from important locations are processed first. In the high-density multilayer circuit board example, the shape of the deep-mining feature set is (1000×1280) (5 locations × 256 dimensions), the same as the fusion-level feature set, but the physical meaning of the features is more explicit (organized by location). This feature set is then finalized to check the completeness of the feature dimensions (whether it covers all locations and dimensions), the range of feature values ​​(whether they are within a reasonable range), and the differences between samples (whether it can distinguish between qualified and unqualified samples). After successful validation, this set is determined as the deep-mining feature set and used as the input for model feature transfer.

[0137] Step S140: Based on the deep mining feature set, adjust the node connection and feature transmission rules inside the deep learning quality inspection analysis model to generate a feature transmission path inside the model that is adapted to the deep mining feature set, and obtain the adapted feature transmission path set of the deep learning quality inspection analysis model.

[0138] The feature propagation module receives the deep-mined feature set as input, analyzes the dimension, distribution, and correlation of the features, and automatically adjusts the node connections and propagation rules within the model. In the example of a high-density multilayer circuit board, the deep-mined feature set has a dimension of 1280, contains features from 5 detection locations, and has complex correlations. The adjustment process is divided into three stages: path evaluation, node reconstruction, and rule optimization. Each stage generates a corresponding adjustment plan and evaluation report to ensure that the propagation path matches the characteristics of the feature set.

[0139] Step S141: Input the deep mining feature set into the feature transmission module of the deep learning quality inspection analysis model, drive the feature transmission module of the deep learning quality inspection analysis model, and calculate the processing efficiency score of the current feature transmission path for the deep mining feature set.

[0140] The feature propagation module incorporates a path evaluator, which calculates three efficiency metrics by simulating the flow of features in the current propagation path: propagation delay (the time it takes for a feature to travel from input to output), feature loss rate (the amount of feature information lost during propagation), and path matching degree (the degree of matching between path nodes and feature dimensions). Each metric is standardized to a score of 0-1, and the processing efficiency score is the weighted sum of the three metrics (weights of 0.3, 0.4, and 0.3, respectively). In the example of a high-density multilayer circuit board, the evaluation results for the current propagation path are: propagation delay 0.72, feature loss rate 0.65, path matching degree 0.58, and a processing efficiency score of 0.63 (below the threshold of 0.8), requiring path adjustment. The evaluation process generates detailed metric calculation logs and bottleneck analysis reports, indicating that the low matching degree is mainly due to the lack of corresponding propagation nodes for the five newly added feature dimensions.

[0141] Step S142: Extract the path representation information of the original feature transmission path of the deep learning quality inspection analysis model. The path representation information of the original feature transmission path of the deep learning quality inspection analysis model includes the node distribution, the connection method between nodes, and the feature transmission rules of the original feature transmission path.

[0142] The path representation information of the original feature propagation path is extracted through the model's structure configuration file and includes three parts: node distribution (node ​​ID, type, function, input / output dimensions), connection method (source node ID, target node ID, connection strength, connection type), and propagation rules (activation function type, gradient clipping threshold, regularization parameters). In the example of a high-density multilayer circuit board, the original path contains 30 propagation nodes (corresponding to 15 preset feature dimensions), 45 node connections (an average of 1.5 connections per node), and 15 propagation rules (1 rule per dimension). The path representation information is stored as a GraphML format file, and the node distribution and connection relationships can be displayed using graph visualization tools.

[0143] Step S143: Compare the feature dimensions and relationships between the feature dimensions of the deep mining feature set with the node distribution and connection rules of the original feature transmission path to generate a list of differences.

[0144] A structured comparison algorithm is used to compare feature dimensions and node distribution (whether the number matches), feature relationships and node connections (whether corresponding connections exist), and feature processing requirements and propagation rules (whether the rules are applicable). In the example of a high-density multilayer circuit board, the difference information list contains three types of differences: node quantity differences (30 nodes vs. 20 feature dimensions, node excess), connection loss differences (5 new feature dimensions have no corresponding connections), and rule mismatch differences (the propagation rules for 3 feature dimensions are not applicable to the new feature type). The difference information list is stored as a CSV file, and each difference record contains four fields: difference type, location, description, and severity (high / medium / low). Differences with high severity need to be processed first.

[0145] Step S144: Adjust the node distribution of the original feature transmission path based on the difference information list, and add or delete transmission nodes according to the number of dimensions of the deep-mined features and the feature correlation to generate the adjusted transmission node distribution.

[0146] For example, step S1441: parse the difference information list, extract the difference content in which the number of transmission nodes reflecting the original feature transmission path does not match the number of dimensions of the deep mining features, and generate node number adaptation difference information.

[0147] Records with a severity level of "High" and a type of "Node Count Difference" are selected from the difference information list. The node ID, corresponding feature dimension, and node function description are extracted. In the example of a high-density multilayer circuit board, the node count adaptation difference information shows that of the original 30 nodes, the feature dimensions corresponding to 10 nodes have been eliminated (no longer included in the expanded feature dimension set), and there are no corresponding nodes for the 5 new feature dimensions. The node count adaptation difference information is organized into a JSON array, with each element containing three fields: node ID, status (excess / missing), and associated feature dimension ID.

[0148] Step S1442: Based on the node number adaptation difference information, calculate the total number of dimensions of the deep mining features, calculate the number of transmission nodes that need to be added or deleted in the original feature transmission path according to the total number of dimensions of the deep mining features, and generate node number adjustment information.

[0149] The total number of dimensions for deep feature mining is 20 (the size of the expanded detection feature dimension set). The original number of transit nodes was 30. Therefore, 10 redundant nodes need to be removed, and 5 new nodes need to be added, resulting in a net reduction of 5 nodes and a adjusted node count of 25. The node count adjustment information includes a list of removed nodes (10 IDs) and requirements for new nodes (5 feature dimension IDs and functional descriptions). The selection criteria for removed nodes are based on their contribution to model performance (calculated through Leave-One-Out cross-validation), with nodes whose contribution is below a threshold being prioritized for removal. The requirements for new nodes specify the node type (convolutional node / recurrent node / graph node) and function (feature enhancement / noise filtering / dimensionality transformation) needed for each new feature dimension.

[0150] Step S1443: If the number of existing transmission nodes is greater than the number of dimensions of the deep mining features, generate new transmission nodes according to the number of missing nodes, assign a basic node identifier to each newly generated transmission node, and generate a set of transmission nodes to be added.

[0151] Based on the requirements for new nodes, five new conduction nodes are generated: Temperature Coefficient of Dielectric Constant Node (Type: Graph Node, Function: Correlation Feature Enhancement), Coefficient of Thermal Expansion Node (Type: Graph Node, Function: Correlation Feature Enhancement), Coating Hardness Node (Type: Convolution Node, Function: Local Feature Extraction), Blind Hole Perpendicularity Node (Type: Convolution Node, Function: Edge Feature Enhancement), and Dielectric Layer Bubble Rate Node (Type: Cyclic Node, Function: Temporal Feature Smoothing). Each new node is assigned a unique node identifier (format "NODE-NEW-XXX") and initial parameters (copied from nodes of the same type and randomly initialized). The set of conduction nodes to be added is stored as an XML configuration file, containing five items: node ID, type, function, input / output dimensions, and initial parameters.

[0152] Step S1444: Extract the dimensional correlation of the deep mining features, and arrange the position of each new transmission node in the set of transmission nodes to be added according to the dimensional correlation of the deep mining features. Place the new transmission nodes between the nodes that match the corresponding feature dimensions in the original transmission path to generate new node layout information.

[0153] Based on the dimensional relationships of the deeply mined features (a set of inter-dimensional relationships from the expanded set of detected feature dimensions), the positions of new transmission nodes are determined. For example, the dielectric constant temperature coefficient node (new) has a causal relationship with the thermal stress distribution node (existing), so it is placed before the thermal stress distribution node; the blind hole perpendicularity node (new) has a cooperative relationship with the hole size node (existing), so it is placed next to the hole size node. The new node layout information includes three items: node ID, coordinate position (x, y), and level (input layer / hidden layer / output layer). The node position is automatically calculated using a spring layout algorithm to ensure that associated nodes are close together. The layout result generates a DOT format file, which can visualize the position of the new node in the path.

[0154] Step S1445: If the number of existing transmission nodes is greater than the number of dimensions of the deep mining features, identify transmission nodes in the original transmission path that are not related to the dimensions of the deep mining features, and generate a set of transmission nodes to be deleted. The transmission nodes in the set of transmission nodes to be deleted do not participate in the transmission and processing of the deep mining features.

[0155] Nodes unrelated to any of the 20 dimensions of the deep-mining features were selected from the existing transmission nodes (by checking if the node's input feature dimension is in the deep-mining feature dimension list). In the example of a high-density multilayer circuit board, 10 nodes were identified to be removed. These nodes correspond to obsolete old feature dimensions (such as the "linewidth deviation rate" dimension) and will not participate in the transmission and processing of new features. The set of transmission nodes to be removed contains three items: node ID, associated old feature dimension, and last used timestamp, arranged in descending order of last used timestamp (the most recently used nodes are removed last for easy rollback).

[0156] Step S1446: Remove the transmission nodes in the set of transmission nodes to be deleted from the original feature transmission path, delete the connection relationships between nodes related to the transmission nodes, and generate the basic path information after the nodes are removed.

[0157] The graph manipulation API is used to delete the node to be deleted and all its associated connections (input and output connections). The deletion process consists of two steps: first, deleting the connections associated with the node to be deleted (removing the record containing the node ID from the connection list), and then deleting the node itself (removing the node ID from the node list). In the example of a high-density multilayer board, after deleting 10 nodes and 15 connections, the path base information contains 20 original nodes and 30 connections. The path base information after node removal is stored as an updated GraphML file, while the original file is backed up for rollback in case of problems.

[0158] Step S1447: Extract the transmission rules of deep mining features, and optimize the node positions of the feature transmission paths after adding new nodes or removing useless nodes according to the transmission rules of deep mining features. Adjust the relative positions of all transmission nodes and generate node position optimization information.

[0159] The deep mining of feature propagation patterns includes: electrical features are propagated first (chronological order), structural features are propagated in parallel (spatial distribution), and material features are fused last (comprehensive judgment). Based on these patterns, a force-directed placement algorithm is used to optimize node positions overall, clustering nodes with similar features and shortening propagation paths. In the example of a high-density multilayer circuit board, after optimization, electrical performance nodes are distributed on the left (x < 0.3), physical structure nodes are distributed on the right (x > 0.7), and material property nodes are distributed in the middle (0.3 ≤ x ≤ 0.7), reducing inter-node connections by 60%. The node position optimization information includes the new coordinates (x, y) of all nodes and the rate of change in path length before and after optimization, stored as a JSON file.

[0160] Step S1448: Integrate the node position optimization information into the adjusted transmission path to generate a preliminary transmission node distribution. The preliminary transmission node distribution includes all retained original nodes, newly added nodes, and optimized node positions.

[0161] The 5 newly added nodes (located from new node layout information) and the 20 retained original nodes (located from node position optimization information) are integrated to generate a preliminary distribution of 25 nodes. Each node's attributes include ID, type, function, coordinates, and input / output dimensions. Nodes are categorized by function into input nodes (5), hidden nodes (15), and output nodes (5). The preliminary distribution of nodes is described using a GraphML file, and the overall layout can be viewed using graph visualization tools to check for overlapping nodes or unreasonable positional distributions. In the example of a high-density multilayer circuit board, the average distance between nodes in the preliminary distribution is 0.15 (normalized coordinates), with no node overlap, and the total path length is reduced by 25% compared to before optimization.

[0162] Step S1449: Extract the efficiency requirements for feature transmission, and normalize the distance between nodes in the preliminary transmission node distribution according to the efficiency requirements for feature transmission. Adjust the transmission distance between adjacent transmission nodes to generate an adjusted transmission node distribution. The number of nodes and their positions in the adjusted transmission node distribution are correlated with the number of dimensions and the correlation relationship of the deep mining features.

[0163] The efficiency requirements for feature propagation include: maximum propagation latency (<10ms), minimum connection strength (>0.5), and average path length (<5 nodes). Based on these requirements, the distances between adjacent nodes in the initial propagation node distribution are normalized: the distances between nodes on critical paths (such as the main path from input to output) are shortened, and the distances on non-critical paths (such as auxiliary feature paths) are increased. In the example of a high-density multilayer circuit board, the adjusted propagation node distribution contains 25 nodes (20 original nodes retained + 5 newly added). After normalization of node coordinates, the average distance on critical paths is shortened by 15%, and the distance on non-critical paths is increased by 10%, meeting the efficiency requirements. The final adjusted propagation node distribution is stored as a GraphML file and a node attribute table (CSV) as the basis for subsequent connection reconstruction.

[0164] Step S145: Configure a feature processing function for each transmission node in the adjusted transmission node distribution. The feature processing function is determined according to the parsing task of the deep mining feature set, and a transmission node function configuration set is generated.

[0165] Based on the analysis tasks of deep feature mining (classification task: qualified / unqualified; regression task: defect probability prediction), feature processing functions are configured for each transmission node. The types of processing functions include: convolutional processing functions (for image features), recurrent processing functions (for temporal features), graph processing functions (for associated features), and fully connected processing functions (for comprehensive features). In the example of a high-density multilayer circuit board, feature preprocessing functions (standardization, dimensionality reduction) are configured for 5 input nodes, feature transformation functions (convolution, recurrent, graph operations, activation functions) are configured for 15 hidden nodes, and feature aggregation functions (pooling, weighted summation) are configured for 5 output nodes. The parameters of each processing function (such as convolution kernel size, number of recurrent units) are determined according to the input and output dimensions of the node. The configuration information is stored in a JSON file, containing four items: node ID, function type, parameter list, and hyperparameter range.

[0166] Step S146: Based on the redefined set of transmission node functions, reconstruct the connection methods between the adjusted transmission nodes, build node connection relationships that conform to the deep mining feature transmission rules, and generate the reconstructed node connection methods.

[0167] Based on the transmission rules of deep mining features (electrical → structural → material → synthesis) and node functions, the connection methods between nodes are reconstructed as follows: input node → hidden node (same type feature) → hidden node (cross type feature) → output node. The connection strength is set according to the strength of the feature association (higher association strength results in greater connection strength). Connection types are divided into necessary connections (strong association, cannot be deleted) and optional connections (weak association, can be dynamically adjusted). In the example of a high-density multilayer circuit board, the reconstructed node connection method contains 55 connections (10 input nodes + 35 hidden nodes + 10 output nodes), of which 30 are necessary connections (55%) and 25 are optional connections (45%). The connection method is stored as an edge list (CSV file), containing five items: source node ID, target node ID, connection strength, connection type, and initial weight value.

[0168] Step S147: Formulate new feature transfer rules based on the reconstructed node connection method. The new feature transfer rules match the function of each transmission node and the connection relationship between nodes to generate the matched feature transfer rules.

[0169] Feature propagation rules include: activation rules (when nodes are activated to process features), gradient rules (how gradients are backpropagated), regularization rules (how to prevent overfitting), and dynamic adjustment rules (how connection strength is updated during training). The activation rule for each node sets a threshold based on the confidence level of the input features (e.g., activation occurs when confidence > 0.6). The gradient rule uses gradient clipping (the threshold is set according to the node type), the regularization rule uses L2 regularization (coefficient 0.001), and the dynamic adjustment rule uses the Hebbian learning rule (connection strength increases with the number of co-activations). In the high-density multilayer circuit board example, the matched feature propagation rules contain 25 node rules (one for each node) and 55 connection rules (one for each connection), stored as an XML rule file. The model loads and parses these rules at runtime.

[0170] Step S148: Integrate the new feature transmission rules into the reconstructed node connection method and the adjusted transmission node distribution to generate a preliminary adapted feature transmission path. The preliminary adapted feature transmission path undertakes the basic transmission and processing of deep mining features.

[0171] The distribution of transmission nodes (GraphML), node connection methods (edge ​​list CSV), and feature transfer rules (XML) are integrated to generate a preliminary adapted feature transfer path. This integration process is implemented using a model compilation tool, converting the three parts of information into a computational graph executable by the model (containing the specific implementations of nodes, edges, and rules). In the high-density multilayer circuit board example, the computational graph of the preliminary adapted feature transfer path contains 25 nodes, 55 edges, and 80 rules. The data flow between nodes and the order of rule application can be visualized using TensorBoard. The preliminary adapted feature transfer path is saved as a model checkpoint file, containing the computational graph structure and initial parameters, and can be directly used for inference or further training.

[0172] Step S149: Input the deep mining feature set into the preliminary adapted feature transmission path to simulate feature transmission and generate the simulated feature output result.

[0173] Simulated transfer was performed using 10% of the samples (100 PCB samples) from the deep-mined feature set. The input and output features, activation state, and processing time of each node were recorded, generating a simulation transfer log. The feature output after simulation transfer includes the predicted value of the output node (pass / fail probability) and the feature transfer trajectory (the flow path of the feature between nodes). In the example of a high-density multilayer PCB, the prediction accuracy of the simulated output was 78% (preliminary adaptation, not trained), the average feature transfer latency was 8ms (meeting the <10ms requirement), and the feature loss rate was 0.15 (better than the threshold of 0.2). The simulation results are stored as HDF5 files, containing prediction results and node state data, for subsequent path optimization.

[0174] Step S1410: Integrate the feature output results after simulation with the feature representation requirements of the deep mining feature set, optimize the details of the initially adapted feature transmission path, generate a fully adapted feature transmission path, and determine the fully adapted feature transmission path as the adapted feature transmission path set of the deep learning quality inspection analysis model.

[0175] Feature representation requirements include prediction accuracy (>95%), feature loss rate (<0.1), and propagation latency (<8ms). By comparing the simulation output with the representation requirements, the path details requiring optimization were identified: adjusting the activation functions of the 5 hidden nodes (from ReLU to LeakyReLU), adding 3 optional connections (to enhance cross-type feature propagation), and optimizing the initial weight values ​​of the 10 connections (adjusted according to the gradient magnitude in the simulation). The optimized fully adapted feature propagation path was simulated again, improving the prediction accuracy to 85%, reducing the feature loss rate to 0.08, and reducing the propagation latency to 7.5ms, meeting all representation requirements. This path was determined as the adapted feature propagation path set, including the final computation graph (model file), optimization log, and performance report—three files in total—serving as the final propagation path for model inference.

[0176] Step S150: Perform full-dimensional feature parsing processing on the deep mining feature set based on the adapted feature transmission path set to generate the circuit board quality inspection test data analysis results corresponding to the circuit board quality inspection test data set.

[0177] After adaptation, the feature propagation path set is loaded with the deep-mined feature set (1000 samples) for full-dimensional feature analysis. The analysis process is divided into three stages: feature input, path propagation, and result output. In the feature input stage, the deep-mined features are converted into model input tensors. In the path propagation stage, features are passed according to the adapted path, and node processing functions and propagation rules are executed. In the result output stage, the features of the output nodes are aggregated to generate the final quality inspection analysis results. In the example of high-density multilayer circuit boards, the full-dimensional feature analysis process takes about 10 minutes (1000 samples) and generates analysis results that include the quality inspection conclusion (pass / fail), defect type (such as blind via deviation, short circuit), defect probability (0-1), and key influencing features (Top 5 contribution) for each sample.

[0178] Step S151: Input the complete set of deep mining features into the set of adapted feature transmission paths, drive the set of adapted feature transmission paths to transmit features along the entire path according to its own transmission path and transmission rules, and generate a full-dimensional feature set after feature transmission.

[0179] The deep-mined feature set (1000×1280) is converted into tensor format (float32) through the model's input layer. After data normalization (Z-score standardization), it is input into the adapted feature propagation path. Features are passed according to the node order in the path: input nodes (5) → hidden nodes (15) → output nodes (5). Each node processes the feature according to the configured processing function and propagation rules. During the propagation process, the strength of optional connections is dynamically adjusted (based on real-time feature confidence) to ensure that important features are passed first. The full-dimensional feature set after feature propagation contains the output features of each node (25 nodes × 1000 samples × average 64 dimensions = 1,600,000 feature values), stored as an HDF5 file for subsequent result parsing and model interpretation.

[0180] Step S152: Extract the feature representation information of each detection part and each detection dimension from the full-dimensional feature set after feature transfer, and split the feature representation information according to the analysis dimensions of circuit board quality inspection and testing to generate multiple analysis dimension feature subsets.

[0181] The analysis dimensions are divided into five categories according to the testing standards: continuity reliability (7 feature dimensions), insulation performance (5 dimensions), structural integrity (4 dimensions), material compatibility (3 dimensions), and process stability (1 dimension). Feature masking technology (defining a binary mask for each analysis dimension) is used to extract the feature representation information of the corresponding dimension from the full-dimensional feature set. In the example of a high-density multilayer circuit board, five feature subsets of analysis dimensions are generated. The feature dimension of each subset is the number of feature dimensions included in that analysis dimension multiplied by the node output dimension, with a shape of approximately (1000×384). The splitting process is implemented through matrix multiplication, multiplying the full-dimensional feature matrix with the analysis dimension mask matrix to extract relevant features.

[0182] Step S153: Input each feature subset of the analysis dimension into a fully connected network layer, and output the quality inspection status score vector of the corresponding analysis dimension as the key quality inspection information of the corresponding analysis dimension.

[0183] Each feature subset of the analysis dimension corresponds to an independent fully connected network layer (expert network), containing two hidden layers (128→64 neurons) and one output layer (1 neuron, sigmoid activation). The fully connected network layer performs a non-linear transformation on the input feature subset of the analysis dimension, outputting a quality inspection status score of 0-1 (1 represents completely acceptable, 0 represents completely unacceptable). In the example of high-density multilayer circuit boards, the quality inspection status score vectors for the five analysis dimensions are all of shape (1000×1), and the vector value represents the pass probability of that dimension. The parameters of each fully connected network layer are jointly optimized during the model training phase to ensure the consistency of scores across dimensions. Key quality inspection information, including the score vector and score confidence (calculated through dropout sampling), is stored as a CSV file.

[0184] Step S154: Compare the quality inspection status score vector of each analysis dimension with the preset pass threshold vector, and generate a binary quality inspection status label based on the comparison result as the quality inspection status representation information of the corresponding analysis dimension.

[0185] The preset pass / fail threshold vector is determined based on historical quality inspection data and industry standards. The thresholds for the five analysis dimensions are: conductivity reliability 0.85, insulation performance 0.90, structural integrity 0.80, material compatibility 0.85, and process stability 0.95. The comparison rule is: score ≥ threshold → label 1 (pass), score < threshold → label 0 (fail). In the example of high-density multilayer circuit boards, each analysis dimension generates a (1000×1) binary label vector, generating a total of 5000 labels across the five dimensions. The quality inspection status representation information includes label vectors, thresholds, comparison rules, and anomaly markers (samples with scores close to the threshold), stored as a JSON file. Anomaly marker samples (scores within ±0.05 of the threshold) are used for manual review.

[0186] Step S155: Summarize the quality inspection status characterization information of all analysis dimensions according to the detection parts of the circuit board, and generate comprehensive quality inspection status characterization information for each detection part. The comprehensive quality inspection status characterization information of each detection part covers the quality inspection characterization results of all analysis dimensions of the part.

[0187] The quality inspection status characterization information is summarized and analyzed across five inspection areas (pads, traces, blind vias, dielectric layers, and edges). Each area includes the label and score for the relevant analysis dimensions. For example, the blind via area involves two analysis dimensions: structural integrity and process stability, summarizing both labels and scores for that area. In the high-density multilayer circuit board example, the comprehensive quality inspection status characterization information for each of the five inspection areas includes results for 2-3 analysis dimensions, stored as a nested JSON structure. The first-level key is the area ID, the second-level key is the analysis dimension ID, and the values ​​are the label and score. The summarization process is implemented using Python dictionary operations, aggregating the analysis dimension results by area ID.

[0188] Step S156: Perform logical operations on the quality inspection status labels of all analytical dimensions in the comprehensive quality inspection status characterization information of each inspection site, and use the operation results as the overall quality inspection pass labels of that inspection site to generate the overall quality inspection characterization content of each inspection site.

[0189] Logical operations use an AND operation: all analysis dimension labels are 1 → overall label 1 (pass); any analysis dimension label is 0 → overall label 0 (fail). For example, if the structural integrity label of a blind via is 0 and the process stability label is 1, then the overall label is 0. The overall quality inspection characterization of each inspected part includes the overall label, a list of failed analysis dimensions, and key defect features (the top 3 features that caused the failure). In the example of a high-density multilayer circuit board, the overall quality inspection characterization of the 5 parts is stored as a JSON array, with each element containing 4 items: part ID, overall label, list of defect dimensions, and list of key features.

[0190] Step S157: Extract the structural sequence of the circuit board, arrange the overall quality inspection characterization content of all inspection parts according to the structural sequence of the circuit board, and generate an ordered overall quality inspection characterization sequence of the circuit board. The ordered overall quality inspection characterization sequence of the circuit board reflects the distribution of the quality inspection status of each part of the circuit board.

[0191] The structural sequence of the circuit board is: inner layer → outer layer → blind via → pad → edge (from core to periphery). The overall quality inspection characterization content of the five inspection locations is arranged according to this order, generating an ordered overall quality inspection characterization sequence for the circuit board. Each element in the sequence retains the overall label and key defect information of the location, facilitating the observation of defect distribution patterns. In the example of a high-density multilayer circuit board, the ordered sequence is stored as a JSON array, with the order being inner layer traces → outer layer traces → blind vias → pads → edge, and the overall label and defect information of each location arranged sequentially.

[0192] Step S158: Count the overall quality inspection pass tags of all detection parts in the ordered overall quality inspection characterization sequence of the circuit board. If all tags are pass, generate an overall pass tag; otherwise, generate an overall fail tag and a list of fail parts, which serve as the core information of the overall quality inspection of the circuit board.

[0193] The algorithm iterates through all the overall labels of the inspected parts in the ordered sequence, performing an AND operation: all 1s indicate an overall pass (label 1), and the presence of 0 indicates an overall fail (label 0). It also records the IDs of all parts with a label of 0. In the example of a high-density multilayer circuit board, out of 1000 samples, 785 samples (78.5%) passed the overall inspection, while 215 samples (21.5%) failed. The main failed parts were blind vias (45%) and outer layer circuitry (30%). The core information for the overall circuit board quality inspection includes the overall label, a list of failed parts, and the defect severity (number of failed parts), stored as a JSON object, serving as the core part of the quality inspection conclusion.

[0194] Step S159: Integrate the overall quality inspection core information of the circuit board with the overall quality inspection characterization content of each test part to generate comprehensive circuit board quality inspection test information covering the whole and the parts. The comprehensive circuit board quality inspection test information covering the whole and the parts includes all quality inspection-related analysis content of the circuit board.

[0195] The fusion process integrates the core information of the overall quality inspection (overall label, failed parts) with the detailed characterization of each part (analysis dimension labels, scores, key features) into a structured document. In the example of high-density multilayer circuit boards, the comprehensive information includes three levels: overall conclusion (1 item) → part details (5 items) → dimension details (up to 3 items / part), forming a tree structure. The fused information is stored in JSON format, facilitating subsequent conversion into report documents or database records. Simultaneously, an information summary (overall label, main defect locations, key impact features) is generated for quick browsing.

[0196] Step S1510: Extract the standard output format of the circuit board quality inspection test data analysis, organize the comprehensive information of the circuit board quality inspection test covering the whole and the parts according to the standard output format of the circuit board quality inspection test data analysis, and generate standardized circuit board quality inspection test data parsing content, which serves as the circuit board quality inspection test data analysis result corresponding to the circuit board quality inspection test data set.

[0197] The standard output format for PCB quality inspection test data analysis follows the IPC-A-600H standard and the company's internal report template, including: basic information (board number, batch, inspection time), overall conclusion (pass / fail), part details (quality inspection results and defect descriptions for each part), defect image links (if any), and improvement suggestions (based on key impact features). Comprehensive information covering both the overall picture and specific parts is organized according to this format, supplementing with basic information (extracted from the original inspection data) and defect image links (queried from the inspection equipment database). In the example of high-density multilayer PCBs, 1000 standardized analysis results are generated (one for each sample), stored as PDF reports (for manual review) and JSON files (for MES system integration). The final PCB quality inspection test data analysis results include the quality inspection conclusions, defect distribution, and improvement suggestions for this batch of 1000 high-density multilayer PCBs.

[0198] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of a deep learning-based circuit board quality inspection and test data analysis system 100 provided in an embodiment of this application. The deep learning-based circuit board quality inspection and test data analysis system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0199] In this embodiment, the machine-readable storage medium 120 can also be integrated into the processor 130 and can communicate and interact with external systems through the communication unit 110. The machine-readable storage medium 120 stores machine-executable instructions for executing the scheme of this application, and the processor 130 executes the machine-executable instructions stored in the machine-readable storage medium 120 to implement the deep learning-based circuit board quality inspection test data analysis method provided in the aforementioned method embodiments.

[0200] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A deep learning-based circuit board quality inspection test data analysis method, characterized by, The method includes: A circuit board quality inspection test data set is obtained. The circuit board quality inspection test data set is mapped and bound to the detection feature dimensions of the deep learning quality inspection analysis model to generate a binding association set of the data dimensions of the circuit board quality inspection test data set and the detection feature dimensions of the deep learning quality inspection analysis model. In this process, the inherent data dimension description information of the circuit board quality inspection test data set is generated by extracting the inherent data dimensions of each type of detection data in the circuit board quality inspection test data set, and the preset detection feature dimensions of the deep learning quality inspection analysis model are extracted to generate the preset detection feature dimension description information of the deep learning quality inspection analysis model. The inherent data dimension description information and the preset detection feature dimension description information are mapped and bound to generate a binding association set. Based on the binding association set, the deep learning quality inspection analysis model is driven to perform dynamic expansion processing of detection feature dimensions, thereby obtaining the expanded detection feature dimension set of the deep learning quality inspection analysis model. Specifically, by parsing the dimension mapping binding information in the binding association set, the inherent data dimensions of the circuit board quality inspection test data set that do not match the preset detection feature dimensions of the deep learning quality inspection analysis model are extracted, generating an unmatched inherent data dimension set. Furthermore, for each inherent data dimension in the unmatched inherent data dimension set, all attribute information of its associated detection data is extracted, generating a refined attribute information set for each unmatched inherent data dimension. The refined attribute information set is then input into the deep learning quality inspection analysis model to construct the expanded detection feature dimension set. The circuit board quality inspection test data set is processed by deep feature mining using the expanded detection feature dimension set to obtain the deep mining feature set of the circuit board quality inspection test data set. Based on the deep mining feature set, the node connection and feature transmission rules inside the deep learning quality inspection analysis model are adjusted to generate a feature transmission path inside the model that is adapted to the deep mining feature set, thereby obtaining the adapted feature transmission path set of the deep learning quality inspection analysis model. Based on the adapted feature transmission path set, the deep mining feature set is subjected to full-dimensional feature parsing processing to generate the circuit board quality inspection test data analysis results corresponding to the circuit board quality inspection test data set.

2. The deep learning-based circuit board quality inspection test data analysis method of claim 1, wherein, The step of acquiring the circuit board quality inspection test data set and mapping and binding the circuit board quality inspection test data set with the detection feature dimensions of the deep learning quality inspection analysis model to generate a binding association set of the data dimensions of the circuit board quality inspection test data set and the detection feature dimensions of the deep learning quality inspection analysis model includes: Extract the inherent data dimensions of each type of test data in the circuit board quality inspection test data set, and generate inherent data dimension description information of the circuit board quality inspection test data set. The inherent data dimension description information of the circuit board quality inspection test data set records the dimension identifier and dimension attribute of each type of test data in the circuit board quality inspection test data set. Extract the preset detection feature dimensions of the deep learning quality inspection analysis model, generate the preset detection feature dimension description information of the deep learning quality inspection analysis model, and the preset detection feature dimension description information of the deep learning quality inspection analysis model records the dimension identifier and dimension attribute of each detection feature dimension in the deep learning quality inspection analysis model. The inherent data dimension description information of the circuit board quality inspection test data set is matched with the preset detection feature dimension description information of the deep learning quality inspection analysis model to generate an initial matching relationship of dimension representation content. Based on the initial matching relationship of the dimensional representation content, each inherent data dimension of the circuit board quality inspection test data set is marked with the corresponding preset detection feature dimension of the deep learning quality inspection analysis model to generate dimension node marking information. The dimension node tagging information is bound to the corresponding inherent data dimension and the corresponding preset detection feature dimension to generate dimension node binding information. Each dimension node binding information corresponds to a set of matched inherent data dimensions and preset detection feature dimensions. All generated dimension node binding information is classified according to the detection data type of the circuit board quality inspection test data set, and a subset of dimension node binding information is generated according to the detection data type. For each subset of dimension node binding information divided by the detection data type, establish the association relationship between the binding information of each dimension node within it, and generate a set of dimension association relationships corresponding to each subset of dimension node binding information divided by the detection data type. Merge the set of dimension relationships corresponding to the subsets of dimension node binding information divided by the data type of detection to generate a unified set of dimension relationships; The integrated set of dimensional relationships is merged with all dimensional node binding information to generate fused dimensional mapping binding information, which includes all matching and association relationships. The fused dimension mapping binding information is determined as the binding association set between the data dimensions of the circuit board quality inspection test data set and the detection feature dimensions of the deep learning quality inspection analysis model. 3.The deep learning-based circuit board quality inspection test data analysis method of claim 1, wherein, The process of dynamically expanding the detection feature dimensions of the deep learning quality inspection analysis model based on the binding association set, resulting in the expanded detection feature dimension set of the deep learning quality inspection analysis model, includes: Parse the dimension mapping binding information in the binding association set, extract the inherent data dimensions of the circuit board quality inspection test data set that do not match the preset detection feature dimensions of the deep learning quality inspection analysis model, and generate a set of unmatched inherent data dimensions; For each inherent data dimension in the set of unmatched inherent data dimensions, extract all attribute information of its associated detection data to generate a refined attribute information set for each unmatched inherent data dimension; The refined attribute information set of each unmatched inherent data dimension is input into the dimension generation module of the deep learning quality inspection analysis model, which drives the dimension generation module of the deep learning quality inspection analysis model to generate new detection feature dimensions corresponding to the unmatched inherent data dimensions. Extract the representation rules of the preset detection feature dimensions of the deep learning quality inspection analysis model, and adjust the representation form of the new detection feature dimension corresponding to the unmatched inherent data dimension according to the representation rules of the preset detection feature dimensions of the deep learning quality inspection analysis model, so as to generate a new detection feature dimension consistent with the preset detection feature dimension representation rules. A dimension identifier is generated for each new detection feature dimension that is consistent with the preset detection feature dimension representation rule. A unique dimension identifier is added to each new detection feature dimension that is consistent with the preset detection feature dimension representation rule. The unique dimension identifier of the new detection feature dimension that is consistent with the preset detection feature dimension representation rule is distinct from the dimension identifier of the preset detection feature dimension and has no duplication. Extract the detection process sequence of the circuit board quality inspection test, and arrange the new detection feature dimension with exclusive dimension identifier and the preset detection feature dimension of the deep learning quality inspection analysis model according to the detection process sequence of the circuit board quality inspection test, so as to generate an ordered sequence of detection feature dimensions. According to the quality inspection analysis logic, a correlation is established between the new detection feature dimension and the preset detection feature dimension in the ordered detection feature dimension sequence, and a set of inter-dimensional correlation relationships is generated. The dimensional correlation between the new detection feature dimension and the preset detection feature dimension in the ordered detection feature dimension sequence is incorporated into the ordered detection feature dimension sequence to generate a detection feature dimension sequence with dimensional correlation. The detection feature dimension sequence with dimensional correlation includes the representation content, identifier and mutual correlation of all dimensions. Based on the set of inter-dimensional relationships, the detection feature dimension sequence with inter-dimensional relationships is topologically sorted to generate a topologically ordered detection feature dimension sequence, which serves as the complete detection feature dimension arrangement result. The complete arrangement of detection feature dimensions is determined as the expanded set of detection feature dimensions of the deep learning quality inspection analysis model.

4. The deep learning-based circuit board quality inspection test data analysis method of claim 1, wherein, The process of performing deep feature mining on the circuit board quality inspection test data set using the expanded detection feature dimension set to obtain a deep-mined feature set for the circuit board quality inspection test data set includes: The expanded set of detection feature dimensions is used as the basis for mining and input into the feature mining module of the deep learning quality inspection analysis model. The feature mining module of the deep learning quality inspection analysis model is set with mining dimension standards that match the circuit board quality inspection test data. The circuit board quality inspection test data set is split according to the type of detection data to generate multiple circuit board quality inspection test data subsets. Each circuit board quality inspection test data subset corresponds to a type of detection feature dimension in the expanded detection feature dimension set. Each subset of circuit board quality inspection test data is matched with the corresponding detection feature dimension for data and dimension adaptation input, and the subset of circuit board quality inspection test data is imported into the mining channel of the corresponding detection feature dimension for feature mining. In each detection feature dimension mining channel, the imported circuit board quality inspection test data subset is processed layer by layer through multiple cascaded feature extraction layers of the deep learning quality inspection analysis model. The output of each layer is used as the input of the next layer to generate a multi-layer feature output set for the corresponding circuit board quality inspection test data subset. Based on a preset feature importance threshold, features in the multi-layer feature output set of each circuit board quality inspection test data subset are filtered, and features with importance higher than the feature importance threshold are retained to generate the core-level feature set of the corresponding circuit board quality inspection test data subset. Arrange the core-level feature set of all circuit board quality inspection test data subsets according to the dimensional order of the expanded detection feature dimension set to generate an ordered core-level feature sequence; The ordered core hierarchical feature sequence is input into the feature fusion layer. The feature fusion layer performs weighted combination and nonlinear transformation on the core hierarchical features from different detection feature dimensions to generate a fused hierarchical feature set. The fused hierarchical feature set is input into the attention mechanism layer, and the weight of each feature is calculated and weighted through the attention mechanism layer to generate a weighted fused hierarchical feature set. The weighted fusion-level feature set is reclassified according to the detection parts of the circuit board quality inspection test, generating fusion-level feature subsets divided by detection parts. Each fusion-level feature subset divided by detection parts corresponds to a specific detection part of the circuit board. All feature subsets divided by detection location are aggregated and integrated to generate a feature set covering all detection locations and all detection dimensions of the circuit board. This feature set covering all detection locations and all detection dimensions of the circuit board is determined as the deep mining feature set of the circuit board quality inspection test data set.

5. The deep learning-based circuit board quality inspection test data analysis method of claim 1, wherein, The process involves adjusting the node connections and feature transfer rules within the deep learning quality inspection analysis model based on the deep mining feature set, generating internal feature transfer paths adapted to the deep mining feature set, and obtaining a set of adapted feature transfer paths for the deep learning quality inspection analysis model, including: The deep-mined feature set is input into the feature transmission module of the deep learning quality inspection and analysis model, which drives the feature transmission module of the deep learning quality inspection and analysis model to calculate the processing efficiency score of the current feature transmission path for the deep-mined feature set. Extract the path representation information of the original feature transmission path of the deep learning quality inspection analysis model. The path representation information of the original feature transmission path of the deep learning quality inspection analysis model includes the node distribution, the connection method between nodes, and the feature transmission rules of the original feature transmission path. The feature dimensions and relationships between the deep-mined feature set are compared with the node distribution and connection rules of the original feature propagation path to generate a list of differences. Based on the list of differences, the node distribution of the original feature transmission path is adjusted. Transmission nodes are added or removed according to the number of dimensions of the deeply mined features and the feature association relationship to generate the adjusted transmission node distribution. Configure a feature processing function for each of the adjusted transmission nodes in the distribution of transmission nodes. The feature processing function is determined according to the parsing task of the deep mining feature set, and a set of transmission node function configurations is generated. Based on the redefined set of transmission node functions, the connection methods between the adjusted transmission nodes are reconstructed to build node connection relationships that conform to the feature transmission rules of deep mining, and generate the reconstructed node connection methods. Based on the reconstructed node connection method, a new feature transfer rule is formulated. The new feature transfer rule matches the function of each transmission node and the connection relationship between nodes to generate a matched feature transfer rule. The new feature transmission rules are integrated into the reconstructed node connection method and the adjusted transmission node distribution to generate a preliminary adapted feature transmission path. The preliminary adapted feature transmission path undertakes the basic transmission and processing of deep mining features. The deep-mined feature set is input into the initially adapted feature transmission path to simulate feature transmission and generate the simulated feature output result. The simulated feature output results are integrated with the feature representation requirements of the deep mining feature set. The details of the initially adapted feature transmission path are optimized to generate a fully adapted feature transmission path. The fully adapted feature transmission path is then determined as the adapted feature transmission path set of the deep learning quality inspection analysis model.

6. The deep learning-based circuit board quality inspection test data analysis method of claim 2, wherein, The step of matching the inherent data dimension description information of the circuit board quality inspection test data set with the preset detection feature dimension description information of the deep learning quality inspection analysis model to generate an initial matching relationship for the dimension representation content includes: The inherent data dimension description information of the circuit board quality inspection test data set is split according to the type of representation content to generate multiple inherent data dimension representation sub-contents, and each inherent data dimension representation sub-content corresponds to a representation attribute of the inherent data dimension. The preset detection feature dimension description information of the deep learning quality inspection analysis model is split according to the same representation content type to generate multiple preset detection feature dimension representation sub-contents. The type of the split preset detection feature dimension representation sub-contents is consistent with the representation content type of the inherent data dimension. Each inherent data dimension representation sub-content is compared one by one with all preset detection feature dimension representation sub-contents to identify sub-content combinations with the same or similar meanings, and generate similar combinations of representation sub-contents. For each similar combination of characterizing sub-content, a similarity representation is performed to generate similarity representation information for the corresponding similar combination of characterizing sub-content. The similarity representation information for the similar combination of characterizing sub-content reflects the degree of similarity and representation association between sub-contents. All generated similar combinations of representation sub-contents are classified according to their inherent data dimension affiliation, and a similar combination subset corresponding to each inherent data dimension is generated. Each similar combination subset corresponding to each inherent data dimension contains all similar combinations of that inherent data dimension and their corresponding similarity representation information. For each subset of similar combinations corresponding to an inherent data dimension, calculate the similarity score of each similar combination within it, select the similar combination with the highest similarity score as the core similar combination of that inherent data dimension, and generate a core similar combination set. Each core similarity combination in the core similarity combination set is associated with its corresponding inherent data dimension and preset detection feature dimension to generate a core similarity combination after association. The core similarity combination after association is associated with the dimension source of each combination. All core similarity combinations after being bound to their respective affiliations are arranged according to the degree of relevance of their representational content to generate an ordered sequence of core similarity combinations. This ordered sequence of core similarity combinations reflects the representational relevance between different dimensions. Based on the detection data type to which the core similarity combination belongs in the ordered core similarity combination sequence, establish logical associations between different types of core similarity combinations and generate a set of logical relationships between combinations; The ordered core similarity combination sequence is fused with the representation association relationship between different core similarity combinations to generate fused dimensional representation matching information, and the fused dimensional representation matching information is determined as the initial matching relationship of the dimensional representation content.

7. The deep learning-based circuit board quality inspection test data analysis method of claim 3, wherein, For each inherent data dimension in the set of unmatched inherent data dimensions, extract all attribute information of its associated detection data to generate a refined attribute information set for each unmatched inherent data dimension, including: Select one inherent data dimension from the set of unmatched inherent data dimensions as the dimension to be processed, extract the original characterization information of the circuit board quality inspection test data corresponding to the dimension to be processed, and obtain the description content of all test data under the dimension to be processed. The original representation information of the dimension to be processed is split according to the acquisition method of the detection data, and multiple sets of original representation sub-information with different acquisition methods are generated. Each set of original representation sub-information corresponds to the representation content under a data acquisition method. For each set of original characterization sub-information collected using different methods, by parsing its data format and content, the parameter items and their values ​​related to circuit board quality inspection recorded therein are identified, and a set of parameter items corresponding to the original characterization sub-information is generated. The data feature set of each set of original representation sub-information is classified according to the type of data feature to generate categorized data feature subsets. Each categorized data feature subset corresponds to a class of data features of the same type. For each subset of data features categorized by type, a corresponding metadata description is generated. The metadata description includes the name, data type, value range, and physical unit of the parameter item, generating metadata description information. The refined feature representation description information corresponding to the original representation sub-information under all acquisition methods is summarized to generate a preliminary refined attribute information set of the dimension to be processed. The preliminary refined attribute information set of the dimension to be processed covers the feature representation of all acquisition methods. Extract the industry standard characterization rules for circuit board quality inspection and testing, and adjust the terminology and form of the preliminary refined attribute information set of the dimension to be processed according to the industry standard characterization rules for circuit board quality inspection and testing to generate an aligned refined attribute information set. For the aligned and refined attribute information set, remove entries that are completely duplicated and entries that are unrelated to the preset list of key dimension attributes to generate a simplified refined attribute information set. The simplified set of detailed attribute information is arranged according to the importance of the data features to generate an ordered set of detailed attribute information, in which the representation content corresponding to the importance of the data features is arranged in order. The ordered set of refined attribute information is determined as the set of refined attribute information for the unmatched inherent data dimension. All dimensions in the set of unmatched inherent data dimensions are processed in sequence to generate a set of refined attribute information for all unmatched inherent data dimensions. 8.The deep learning-based circuit board quality inspection test data analysis method of claim 4, wherein, In each detection feature dimension mining channel, the imported circuit board quality inspection test data subset is processed layer by layer through multiple cascaded feature extraction layers of the deep learning quality inspection analysis model. The output of each layer serves as the input of the next layer, generating a multi-layer feature output set for the corresponding circuit board quality inspection test data subset, including: A subset of circuit board quality inspection test data imported into the detection feature dimension mining channel is used as initial mining data. The surface data representation content of the initial mining data is extracted to generate the surface feature set of the initial mining data. The surface feature set of the initial mining data reflects the intuitive representation attributes of the data. The surface feature set of the initial mining data is input into the first feature extraction layer of the mining channel. The surface features are processed by the first feature extraction layer to mine the shallow hidden features behind the surface features and generate the shallow hidden feature set of the initial mining data. The shallow latent feature set of the initial mining data is passed from the first feature extraction layer of the mining channel to the second feature extraction layer. The shallow latent features are then processed in depth by the second feature extraction layer to mine the middle latent features behind the shallow latent features, thereby generating the middle latent feature set of the initial mining data. The set of middle-level hidden features of the initial mining data is passed from the second feature extraction layer of the mining channel to the third feature extraction layer. The middle-level hidden features are further processed by the third feature extraction layer to mine the deep hidden features behind the middle-level hidden features, and generate the set of deep hidden features of the initial mining data. The deep implicit feature set of the initial mining data is passed from the third feature extraction layer of the mining channel to the subsequent feature extraction layers. Feature mining is carried out in sequence according to the aforementioned feature processing method, gradually delving into the underlying representation of the data. In each subsequent feature extraction layer, at least one of the following operations—convolution, pooling, or fully connected—is performed on the received feature set from the previous layer to generate the output feature set of the corresponding feature extraction layer. Record the mapping relationship between the feature set generated by each feature extraction layer and its input feature set to generate inter-layer feature mapping relationship information. The inter-layer feature mapping relationship information reflects the transmission and evolution relationship between features of different extraction layers. The feature sets generated by all feature extraction layers, including the surface feature set, the hidden feature sets of each layer, and the feature mapping relationship information between layers, are summarized to generate the feature extraction summary information of the initial mining data; The feature extraction summary information of the initial mining data is arranged according to the hierarchical order of feature extraction to generate a multi-layer feature output set, which reflects the progressive extraction process of features from the surface to the bottom layer. 9.The deep learning-based circuit board quality inspection test data analysis method of claim 1, wherein, The step of performing full-dimensional feature parsing processing on the deep-mined feature set based on the adapted feature transmission path set to generate circuit board quality inspection test data analysis results corresponding to the circuit board quality inspection test data set includes: The deep mining feature set is completely input into the adapted feature transmission path set, which drives the adapted feature transmission path set to transmit features along the entire path according to its own transmission path and transmission rules, generating a full-dimensional feature set after feature transmission. Extract the feature representation information of each detection part and each detection dimension from the full-dimensional feature set after feature transfer, and split the feature representation information according to the analysis dimension of circuit board quality inspection and testing to generate multiple analysis dimension feature subsets; Each feature subset of an analysis dimension is input into a fully connected network layer, and the fully connected network layer outputs the quality inspection status score vector of the corresponding analysis dimension, which serves as the key quality inspection information for the corresponding analysis dimension. The quality inspection status score vector of each analysis dimension is compared with the preset pass threshold vector. Based on the comparison result, a binary quality inspection status label is generated as the quality inspection status representation information of the corresponding analysis dimension. The quality inspection status characterization information of all analysis dimensions is summarized according to the inspection parts of the circuit board to generate comprehensive quality inspection status characterization information for each inspection part. The comprehensive quality inspection status characterization information of each inspection part covers the quality inspection characterization results of all analysis dimensions of the part. Logical operations are performed on the quality inspection status labels of all analytical dimensions in the comprehensive quality inspection status characterization information of each inspection site. The operation results are used as the overall quality inspection labels of the inspection site to generate the overall quality inspection characterization content of each inspection site. Extract the structural sequence of the circuit board, arrange the overall quality inspection characterization content of all inspection parts according to the structural sequence of the circuit board, and generate an ordered overall quality inspection characterization sequence of the circuit board. The ordered overall quality inspection characterization sequence of the circuit board reflects the distribution of the quality inspection status of each part of the circuit board. The overall quality inspection pass labels of all test parts in the ordered overall quality inspection characterization sequence of the circuit board are counted. If all labels are pass, an overall pass label is generated; otherwise, an overall fail label and a list of fail parts are generated as the core information of the overall quality inspection of the circuit board. The core information of the overall quality inspection of the circuit board is integrated with the overall quality inspection characteristics of each testing part to generate comprehensive information on the quality inspection of the circuit board covering both the whole and the parts. The comprehensive information on the quality inspection of the circuit board covering both the whole and the parts includes all quality inspection-related analysis content of the circuit board. Extract the standard output format of the circuit board quality inspection and testing data analysis, and organize the comprehensive information of circuit board quality inspection and testing covering the whole and the parts according to the standard output format of the circuit board quality inspection and testing data analysis to generate standardized circuit board quality inspection and testing data parsing content, which serves as the circuit board quality inspection and testing data analysis result corresponding to the circuit board quality inspection and testing data set.

10. A circuit board quality inspection and testing data analysis system based on deep learning, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the deep learning-based circuit board quality inspection test data analysis method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images