A PCB parameter retrieval method and system based on multi-modal semantic verification and information density filtering

Abstract

The application discloses a PCB parameter retrieval method and system based on multi-modal semantic verification and information density filtering, solves the technical problem that the traditional PCB parameter retrieval method destroys the basic data conditions of parameter analysis, thereby causing the distortion of the PCB parameter extraction result. The method comprises the following steps: acquiring PCB multi-source heterogeneous document data and carrying out multi-modal layout analysis to generate modal slice data; using a pre-trained OCR model and a pre-trained VLM model to complete parallel perception extraction, obtaining OCR text sequence data and document global logical architecture data with constraint text semantic consistency; completing physical logic rectification with the help of a MAP model; synthesizing structured PCB parameter basic data through a parameter key value association algorithm; when responding to a retrieval request, filtering noise by relying on a vector retrieval model, and finally outputting high-precision PCB design parameter retrieval data through fine rearrangement processing.

Description

Technical Field

[0001] This invention relates to the field of printed circuit board design and manufacturing technology, and in particular to a PCB parameter retrieval method and system based on multimodal semantic verification and information density filtering. Background Technology

[0002] In the field of printed circuit board (PCB) design and manufacturing, the industry as a whole is characterized by a complex documentation system and stringent design parameter requirements. Engineers need to rely on a massive amount of industry specifications, internal enterprise standards, and board-level documentation (such as Gerber file notes) to carry out their work. The overall information processing pressure is extremely high, and accurately completing the entry of hundreds of design parameters is the core prerequisite for ensuring the compliant and efficient progress of PCB production and manufacturing.

[0003] Based on the aforementioned industry work scenarios, PCB parameter retrieval, as a core support link for design parameter filling, suffers from prominent problems such as low information retrieval efficiency and easy deviation in content understanding. This shortcoming not only makes the parameter filling process time-consuming and labor-intensive, but also makes it easy to make mistakes in parameter filling due to incomplete information retrieval and misunderstanding, ultimately leading to increased PCB manufacturing costs and project delays, becoming a key pain point restricting the efficiency and quality of PCB design and manufacturing.

[0004] Traditional PCB parameter retrieval methods, when handling minute PCB-specific annotations, Gerber document notes, and high-density specification tables, are prone to OCR errors such as misidentification of key characters due to a lack of PCB-specific domain knowledge constraints. This leads to physical and logical errors such as misaligned rows and columns in the table and mismatched annotations. Furthermore, traditional retrieval mechanisms lack efficient noise reduction capabilities, introducing a large amount of invalid noise during the retrieval stage, directly resulting in inaccurate target parameter location. These two types of underlying data defects propagate to downstream processes, compromising the fundamental data conditions for parameter parsing and causing distortion in the extracted PCB parameters. Summary of the Invention

[0005] This invention provides a PCB parameter retrieval method and system based on multimodal semantic verification and information density filtering, which solves the technical problem that traditional PCB parameter retrieval methods destroy the basic data conditions for parameter parsing, thus causing distortion of PCB parameter extraction results.

[0006] The first aspect of this invention provides a PCB parameter retrieval method based on multimodal semantic verification and information density filtering, comprising:

[0007] Acquire PCB multi-source heterogeneous document data, perform multimodal layout analysis on the PCB multi-source heterogeneous document data, and output modal slice data;

[0008] The modal slice data is extracted in parallel using a pre-trained OCR model and a pre-trained VLM model, and the output is OCR text sequence data and document global logical architecture data used to constrain the semantic consistency of the text sequence.

[0009] The MAP model is used to perform physical and logical correction on the OCR text sequence data based on the prior knowledge base of the PCB domain and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and the corrected text sequence data is output.

[0010] A parameter key-value association algorithm is used to perform structured data synthesis on the corrected text sequence data and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and the structured PCB parameter basic data is output.

[0011] When a user PCB parameter retrieval request is received, a vector retrieval model is used to filter noise from the structured PCB parameter base data and the user PCB parameter retrieval request, and output high-quality document fragment data.

[0012] The high-quality document fragment data is finely rearranged to output high-precision PCB design parameter retrieval data.

[0013] Optionally, the step of performing multimodal layout analysis on the PCB multi-source heterogeneous document data and outputting modal slice data includes:

[0014] Region recognition is performed on the PCB multi-source heterogeneous document data to obtain text area, parameter table area and Gerber visual annotation area;

[0015] Based on the text area, the parameter table area, and the Gerber visual annotation area, the PCB multi-source heterogeneous document data is divided into data segments to obtain the corresponding region data for the text area, parameter table area, and Gerber visual annotation area.

[0016] By integrating the region data corresponding to the text area, the parameter table area, and the Gerber visual annotation area, modal slice data is obtained.

[0017] Optionally, the step of using a pre-trained OCR model and a pre-trained VLM model to perform parallel perceptual extraction on the modal slice data, outputting OCR text sequence data and document global logical architecture data for constraining the semantic consistency of the text sequence, includes:

[0018] The pre-trained OCR model is used to perform character-level scanning and extraction on the modal slice data to obtain the original OCR text sequence data;

[0019] The original OCR text sequence data is input into the pre-trained VLM model for review;

[0020] If the review fails, the pre-trained VLM model is used to correct the original OCR text sequence data to obtain OCR text sequence data.

[0021] If the review is approved, the original OCR text sequence data will be used as the OCR text sequence data.

[0022] The pre-trained VLM model is used to perform spatial logic perception processing on the modal slice data to obtain the table header membership relationship and the leader pointing information;

[0023] By integrating the table header membership relationship and the leader pointing information, a document global logical architecture data is obtained to constrain the semantic consistency of the text sequence.

[0024] Optionally, the step of using the MAP model to perform physical logic correction on the OCR text sequence data based on the PCB domain prior knowledge base and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and outputting the corrected text sequence data, includes:

[0025] The MAP model is used to solve the optimal text sequence problem on the OCR text sequence data based on the document global logical architecture data used to constrain the semantic consistency of the text sequence and the PCB domain prior knowledge base, so as to obtain the solved text sequence.

[0026] Based on the prior knowledge base in the PCB field, character error correction is performed on the solved text sequence to obtain the error-corrected text sequence.

[0027] The document global logical architecture data used to constrain the semantic consistency of the text sequence is used to perform table row and column alignment on the corrected text sequence to obtain the corrected text sequence data.

[0028] Optionally, the step of using a vector retrieval model to filter noise from the structured PCB parameter base data and the user PCB parameter retrieval request, and outputting high-quality document fragment data, includes:

[0029] The vector retrieval model is used to perform vector similarity matching between the structured PCB parameter base data and the user's PCB parameter retrieval request, resulting in multiple candidate document fragments.

[0030] Information density scores are performed on multiple candidate document fragments to obtain an information density score for each candidate document fragment.

[0031] The information density score of each candidate document fragment is compared with a preset dynamic scoring threshold.

[0032] Any candidate document fragment data corresponding to an information density score greater than or equal to the preset dynamic scoring threshold is used as high-quality document fragment data.

[0033] A second aspect of this invention provides a PCB parameter retrieval system based on multimodal semantic verification and information density filtering, comprising:

[0034] The acquisition module is used to acquire PCB multi-source heterogeneous document data, perform multimodal layout analysis on the PCB multi-source heterogeneous document data, and output modal slice data.

[0035] The extraction module is used to perform parallel perceptual extraction of the modal slice data using a pre-trained OCR model and a pre-trained VLM model, and output OCR text sequence data and document global logical architecture data for constraining the semantic consistency of the text sequence.

[0036] The correction module is used to perform physical and logical correction on the OCR text sequence data using the MAP model based on the PCB domain prior knowledge base and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and output the corrected text sequence data.

[0037] The synthesis module is used to perform structured data synthesis on the corrected text sequence data and the document global logical architecture data used to constrain the semantic consistency of the text sequence using a parameter key-value association algorithm, and output structured PCB parameter basic data.

[0038] The filtering module is used to filter noise from the structured PCB parameter base data and the user PCB parameter retrieval request using a vector retrieval model when a user PCB parameter retrieval request is received, and output high-quality document fragment data.

[0039] The rearrangement module is used to perform fine-grained rearrangement of the high-quality document fragment data and output high-precision PCB design parameter retrieval data.

[0040] Optionally, the acquisition module is specifically used for:

[0041] Region recognition is performed on the PCB multi-source heterogeneous document data to obtain text area, parameter table area and Gerber visual annotation area;

[0042] Based on the text area, the parameter table area, and the Gerber visual annotation area, the PCB multi-source heterogeneous document data is divided into data segments to obtain the corresponding region data for the text area, parameter table area, and Gerber visual annotation area.

[0043] By integrating the region data corresponding to the text area, the parameter table area, and the Gerber visual annotation area, modal slice data is obtained.

[0044] A third aspect of the present invention provides an electronic device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the PCB parameter retrieval method based on multimodal semantic verification and information density filtering as described above.

[0045] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the PCB parameter retrieval method based on multimodal semantic verification and information density filtering as described above.

[0046] The fifth aspect of the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the steps of the PCB parameter retrieval method based on multimodal semantic verification and information density filtering as described above.

[0047] As can be seen from the above technical solutions, the present invention has the following advantages:

[0048] The above-mentioned technical solution of the present invention provides a PCB parameter retrieval method based on multimodal semantic verification and information density filtering. It acquires multi-source heterogeneous PCB document data, performs multimodal layout analysis on the multi-source heterogeneous PCB document data, and outputs modal slice data. A pre-trained OCR model and a pre-trained VLM model are used to perform parallel perceptual extraction on the modal slice data, outputting OCR text sequence data and document global logical architecture data for constraining the semantic consistency of the text sequence. A MAP model is used to perform physical logical correction on the OCR text sequence data based on the PCB domain prior knowledge base and the document global logical architecture data for constraining the semantic consistency of the text sequence, outputting corrected text sequence data. A parameter key-value association algorithm is used to perform structured data synthesis on the corrected text sequence data and the document global logical architecture data for constraining the semantic consistency of the text sequence, outputting structured PCB parameter basic data. When a user's PCB parameter retrieval request is received, a vector retrieval model is used to retrieve the structured PCB parameter basic data and the user's P... The CB parameter retrieval request undergoes noise filtering to output high-quality document fragment data. This high-quality document fragment data is then finely rearranged to output high-precision PCB design parameter retrieval data. Based on this scheme, this invention relies on multimodal layout analysis to provide regularized slice data for subsequent data processing. It extracts both text characters and document spatial logic information in parallel using OCR and VLM, and then uses a MAP model combined with domain prior knowledge and a global logical architecture to complete physical logic correction. This corrects character misidentification and layout misalignment problems caused by the lack of domain constraints in traditional methods from the source, eliminating erroneous underlying data caused by source document recognition bias. This lays a solid foundation for accurate data extraction. Combined with a parameter key-value association algorithm, it achieves structured data synthesis, further ensuring data standardization and effectiveness. Subsequently, a vector retrieval model filters noise, and by quantifying the physical parameter value and information entropy of text fragments, low-value noise data is dynamically removed during the retrieval stage. This effectively solves the problem of low effective parameter density in fragments, which leads to problems with large models (LLM). The problem of language models being misled by noise, resulting in illusions or omissions of key parameters is addressed through refined rearrangement and optimization, which effectively ensures the accuracy and reliability of PCB design parameter extraction results. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1This is a flowchart illustrating the steps of a PCB parameter retrieval method based on multimodal semantic verification and information density filtering, as provided in Embodiment 1 of the present invention.

[0051] Figure 2 The process for precision text extraction and semantic verification provided in Embodiment 1 of the present invention;

[0052] Figure 3 This invention provides a method for filtering source document noise using Information Density Score (IDS) as described in Embodiment 1 of the present invention.

[0053] Figure 4 This is a structural block diagram of a PCB parameter retrieval system based on multimodal semantic verification and information density filtering, provided in Embodiment 2 of the present invention. Detailed Implementation

[0054] This invention provides a PCB parameter retrieval method and system based on multimodal semantic verification and information density filtering, which solves the technical problem that traditional PCB parameter retrieval methods destroy the basic data conditions for parameter parsing, thus causing distortion of PCB parameter extraction results.

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that in the optional embodiments of the present invention, the object information and other related data involved require the permission or consent of the object when the embodiments of the present invention are applied to specific products or technologies, and the collection, use, and processing of related data must comply with relevant laws, regulations, and standards. That is to say, if the embodiments of the present invention involve data related to the object, it needs to be obtained with the authorization and consent of the object, the authorization and consent of relevant departments, and in compliance with relevant laws, regulations, and standards. If personal information is involved in the embodiments, the acquisition of all personal information requires the consent of the individual. If sensitive information is involved, the separate consent of the information subject is required, and the embodiments also need to be implemented with the authorization and consent of the object.

[0056] Terminology Explanation:

[0057] 1. PCB (Printed Circuit Board): An electronic interconnection assembly that uses an insulating substrate as a carrier and forms conductive lines and pads through etching process. It is used to realize the mechanical fixation and electrical connection of electronic components. As the core carrier of electronic systems, its circuit layout directly determines signal integrity, electromagnetic compatibility and heat dissipation efficiency. It is an indispensable basic hardware in various electronic devices.

[0058] 2. RAG (Retrieval-Augmented Generation): A technical architecture that combines information retrieval with large-scale language model generation, aiming to solve the problems of model knowledge lag and illusion. This mechanism retrieves relevant context from an external knowledge base before generating a response and uses it as input to the model, thereby significantly improving the factual accuracy, timeliness, and traceability of the output content.

[0059] 3. VLM (Vision-Language Model): A multimodal artificial intelligence model that can simultaneously understand and process images and text; it achieves visual question answering and reasoning in complex scenes by integrating computer vision and natural language processing capabilities.

[0060] 4. IDS (Information Density Score): In current technical solutions for PCB traceability tasks, it is defined as an evaluation metric that quantifies the "effective knowledge content" of text chunks.

[0061] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a PCB parameter retrieval method based on multimodal semantic verification and information density filtering, as provided in Embodiment 1 of the present invention.

[0062] This invention provides a PCB parameter retrieval method based on multimodal semantic verification and information density filtering, comprising:

[0063] Step 101: Obtain PCB multi-source heterogeneous document data, perform multimodal layout analysis on the PCB multi-source heterogeneous document data, and output modal slice data.

[0064] PCB multi-source heterogeneous document data refers to PCB-related document data from diverse sources and with inconsistent structures and formats, including Gerber file annotations, industry specification documents, enterprise internal standards, and board-level description documents.

[0065] It should be noted that by performing region identification and data segmentation on the original document, standardized processing data adapted to subsequent multimodal extraction is formed, providing a regular processing foundation for the parallel perception extraction process.

[0066] Further, step 101 may include the following sub-steps:

[0067] S11. Perform region recognition on PCB multi-source heterogeneous document data to obtain text area, parameter table area and Gerber visual annotation area.

[0068] S12. Based on the text area, parameter table area, and Gerber visual annotation area, the PCB multi-source heterogeneous document data is divided into data segments to obtain the corresponding region data for the text area, parameter table area, and Gerber visual annotation area.

[0069] S13. Integrate the region data corresponding to the text area, parameter table area, and Gerber visual annotation area to obtain modal slice data.

[0070] The text area refers to the independent content area in a multi-source heterogeneous PCB document that contains text descriptions, specifications, and annotations.

[0071] The parameter table area refers to the structured content area in the PCB multi-source heterogeneous document that lists specifications and process parameters in tabular form.

[0072] The Gerber visual annotation area refers to the visual content area within a multi-source heterogeneous PCB document that is generated from Gerber files and carries graphic annotations and dimensional notes.

[0073] It should be noted that the key to constructing the extraction process lies in establishing a complementary perception mechanism between visual and textual features. This invention divides the document into multiple modal slices and automatically identifies text areas, parameter table areas, and Gerber visual annotation areas in the document. This region division is not only used for data separation but also to provide spatial constraints for subsequent semantic verification, ensuring that the processing process is spatially context-aware.

[0074] Specifically, the process involves region identification of heterogeneous PCB multi-source document data to obtain text areas, parameter table areas, and Gerber visual annotation areas. This process relies on the document's visual features and content layout patterns to accurately locate and distinguish functional areas. Based on the text areas, parameter table areas, and Gerber visual annotation areas, the heterogeneous PCB multi-source document data is directionally split and divided according to the defined regional boundaries. The corresponding regional data of the text areas, parameter table areas, and Gerber visual annotation areas are accurately separated and obtained. Then, the regional data corresponding to the text areas, parameter table areas, and Gerber visual annotation areas are integrated according to the original logical order of the document. Different types of fragmented document content are organized to obtain modal slice data. This processing method can achieve the classification and sorting of differentiated information, avoid the recognition disorder caused by the simultaneous processing of multiple types of mixed data, and alleviate the layout logic misalignment problem that is prone to occur in traditional solutions.

[0075] Step 102: Use a pre-trained OCR model and a pre-trained VLM model to perform parallel perceptual extraction on the modal slice data, and output OCR text sequence data and document global logical architecture data for constraining the semantic consistency of the text sequence.

[0076] A pre-trained OCR model (Pre-trained Optical Character Recognition) is a pre-trained optical character recognition model that has been trained on a large amount of text recognition data and can quickly extract and recognize characters from target text content.

[0077] A pre-trained Visual Language (VLM) model refers to a pre-trained visual language model that has been trained on multimodal data and can perceive the spatial logic and regional relationships of a document.

[0078] OCR text sequence data refers to the final text data after compliance review by a pre-trained VLM model. When the review is successful, the original error-free OCR text sequence data is used directly. When the review fails, the corresponding text content is generated after character correction and logical modification by the pre-trained VLM model. The overall data uniformly meets the character specifications and basic semantic requirements of PCB documents, providing a standard text carrier for subsequent spatial logic parsing and parameter processing.

[0079] The document global logical architecture data used to constrain the semantic consistency of text sequences refers to the set of logical information extracted from modal slice data through a pre-trained VLM model to unify and standardize content logic. It can constrain the semantic consistency of various text sequences throughout the process, specifically including the spatial relationship between text areas and parameter table areas, the correspondence between Gerber files and parameter annotations, the membership logic of each content area, the hierarchical relationship of parameter information, etc. It fully reflects the overall structure and element relationship of PCB-related documents, avoids text semantic conflicts and layout logic disorder, and provides standardized logical constraints and structured support for subsequent parameter matching, physical logic verification and text correction.

[0080] It should be noted that a pre-trained OCR model and a pre-trained VLM model are used to perform parallel perceptual extraction of modal slice data. The pre-trained OCR model performs character-level scanning extraction of text content in the modal slice data, while the pre-trained VLM model performs logical perception of the spatial arrangement and regional relationships of the slice data. The two run synchronously and cooperate with each other to finally output OCR text sequence data and document global logical architecture data used to constrain the semantic consistency of the text sequence.

[0081] Furthermore, step 102 may include the following sub-steps:

[0082] S21. Use a pre-trained OCR model to perform character-level scanning extraction on the modal slice data to obtain the original OCR text sequence data;

[0083] S22. Input the original OCR text sequence data into the pre-trained VLM model for review;

[0084] S23. If the review fails, the original OCR text sequence data will be corrected using a pre-trained VLM model to obtain the OCR text sequence data.

[0085] S24. If the review is approved, the original OCR text sequence data will be used as the OCR text sequence data.

[0086] S25. Use a pre-trained VLM model to perform spatial logic perception processing on the modal slice data to obtain the table header membership relationship and the leader pointing information;

[0087] S26. Integrate the table header membership and leader line pointing information to obtain the document global logical architecture data used to constrain the semantic consistency of the text sequence.

[0088] The header affiliation refers to the attribution and hierarchical association information between various header items and their corresponding parameter cells in the PCB parameter table. It is the core logical basis for maintaining the correspondence between table parameters.

[0089] Leader line information refers to the positional correspondence and content binding information between leader lines, graphic guides and corresponding parameter annotations, and Gerber additional annotations in the document, which is used to clarify the relationship between graphic and text elements.

[0090] Raw OCR text sequence data refers to the set of text data that has not been corrected and retains the original character information after the modal slice data has been extracted by a pre-trained OCR model through character-level scanning. It is the basis for subsequent data processing.

[0091] It should be noted that general-purpose OCR models often fail to accurately recognize specialized annotations and complex layout relationships in the PCB field (e.g., misinterpreting 13µm as 130µm or causing misalignment of rows and columns in tables). Since the accuracy of parameter extraction for text models is highly dependent on the quality of the input content, this section aims to significantly improve the accuracy of model input through the synergy of OCR and Visual Large Model (VLM). This solution utilizes an OCR model for initial text extraction from PCB documents, supplemented by a VLM model for secondary semantic verification and physical logic correction of the extraction results, thereby providing high-fidelity, structured foundational data for subsequent deep retrieval systems.

[0092] Furthermore, for the OCR path: a high-precision OCR model is used to perform character-level scanning to obtain preliminary text stream data. For the VLM path: the spatial perception capability of the large visual model is used to capture the global logical structure of the document (such as table header relationships and leader lines).

[0093] Specifically, a pre-trained OCR model is used to perform character-level scanning and extraction of modal slice data, accurately capturing various texts, parameter symbols, and table content within the image to obtain raw OCR text sequence data. This raw OCR text sequence data is then input into a pre-trained VLM model for comprehensive content review. This model uses PCB industry terminology standards, character writing specifications, table row and column matching logic, and the correspondence between annotations and drawings as its core judgment criteria. Once issues such as character misidentification / confusion, abnormal professional parameter descriptions, misaligned text and image annotations, or semantic conflicts in table content are detected, the review is deemed unsuccessful. Subsequently, the pre-trained VLM model is used to further address erroneous characters, contradictory semantics, and disordered relationships. The system performs targeted corrections to obtain OCR text sequence data. If the original OCR text sequence data has accurate character recognition, uses standardized professional terminology, matches parameter annotations and chart content, and has no conflicts in overall text semantics and layout logic, it is deemed to have passed the review. The original OCR text sequence data is then directly used as the OCR text sequence data. Based on this, a pre-trained VLM model is used to perform spatial logic perception processing on the modal slice data, analyze the positional associations and subordinate rules of multiple regions in the document, stably obtain the table header affiliation relationship and leader line pointing information, and then summarize, integrate and logically sort out the two types of association information to obtain the document global logical architecture data used to constrain the semantic consistency of the text sequence.

[0094] Step 103: Using the MAP model, the OCR text sequence data is physically and logically corrected based on the prior knowledge base of the PCB domain and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and the corrected text sequence data is output.

[0095] The PCB prior knowledge base refers to a set of knowledge specifically for the PCB field, which includes various specifications, parameter standards, and anomaly handling rules related to PCB design and production, providing a basis for physical logic correction.

[0096] The MAP model (Maximum A Posteriori Model) is a model used to physically and logically correct and rectify data biases in raw text data. This model can rely on prior knowledge of the given domain and global logical constraints of the document to solve for the optimal sequence and correct errors in the recognized text content. It can specifically correct character recognition errors and logical contradictions in professional scenarios.

[0097] It should be noted that the MAP model is used, combined with a prior knowledge base in the PCB field and document global logical architecture data used to constrain the semantic consistency of the text sequence, to perform physical logical verification on the OCR text sequence data. This corrects potential issues such as character misjudgment and logical deviations, and outputs corrected text sequence data, providing accurate data support for subsequent structured processing.

[0098] Furthermore, step 103 may include the following sub-steps:

[0099] S31. Using the MAP model, based on the document global logical architecture data used to constrain the semantic consistency of the text sequence and the PCB domain prior knowledge base, the optimal text sequence is solved for the OCR text sequence data, and the solved text sequence is obtained.

[0100] S32. Based on the prior knowledge base of the PCB field, perform character error correction on the solved text sequence to obtain the error-corrected text sequence.

[0101] S33. Use the document global logical architecture data used to constrain the semantic consistency of the text sequence to perform table row and column alignment on the corrected text sequence to obtain the corrected text sequence data.

[0102] It should be noted that the core logic of using VLM to review the raw data extracted by OCR lies in constructing a closed-loop correction system based on visual facts and industry experience. In order to transform the expert's subjective corrective experience into a quantifiable calculation process, this invention introduces the maximum a posteriori probability estimation (MAP) model to describe the verification logic.

[0103] Furthermore, this invention does not simply replace text, but rather uses a Visual Model (VLM) to solve for the optimal corrected text sequence, i.e., the solved text sequence T*, under a given visual context and industry common sense. Its formal expression is as follows:

[0104] T*=argmaxP(T|I,T ocr ,K pcb );

[0105] Where I represents the input image slice (i.e., modal slice data), and T... ocr For OCR text sequence data, K pcb T represents the pre-set PCB domain prior knowledge base. During the MAP model solution process, given modal slice data, OCR text sequence data, and PCB domain prior knowledge base constraints, any text sequence in the set of all possible candidate corrected text sequences is the candidate object for solving the optimal corrected text sequence.

[0106] Furthermore, physical logic backtracking is performed to address common OCR recognition errors:

[0107] 1) OCR recognition: T_HOLE_COP_MIN130m → VLM correction: Corrected to 13um by combining image features and PCB manufacturing knowledge).

[0108] 2) OCR recognition: complex table row and column serialization → VLM correction: using visual alignment capabilities to restore the correct key-value pair relationship.

[0109] In this embodiment, the MAP model is adopted, combined with the layout hierarchy and regional association information carried by the document global logical architecture data used to constrain the semantic consistency of the text sequence. At the same time, relying on the industry standards and professional rules in the PCB field prior knowledge base, the optimal text sequence solution is performed on the OCR text sequence data. By comprehensively considering visual context constraints and industry common sense, the candidate text sequence that best fits the PCB document scenario logic is selected, and the solved text sequence is obtained. Then, based on the preset parameter definitions, unit identifiers, common terms and other verification rules in the PCB field prior knowledge base, the solved text sequence is subjected to character-level verification and error correction, correcting character misrecognition and identifier confusion problems that do not conform to professional standards, and obtaining the corrected text sequence. Finally, according to the table header affiliation and row and column correspondence information recorded in the document global logical architecture data used to constrain the semantic consistency of the text sequence, the corrected text sequence is subjected to table row and column alignment processing to make the affiliation relationship between the text content and the table structure accurately match, and obtain the corrected text sequence data.

[0110] Step 104: Use the parameter key-value association algorithm to perform structured data synthesis on the corrected text sequence data and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and output the basic data of structured PCB parameters.

[0111] The parameter key-value association algorithm relies on the layout association rules and spatial constraints in the global logical architecture data of the document, which are used to constrain the semantic consistency of the text sequence. Combining these spatial constraints, it performs targeted matching of parameter items (keys) and corresponding values ​​(values) in the corrected text sequence data. Simultaneously, it verifies the consistency and resolves conflicts of candidate key-value pairs, eliminating invalid or misaligned key-value combinations. This ensures that parameter items and their corresponding values ​​correspond accurately and logically without contradictions, ultimately establishing a standardized key-value correspondence. This provides standardized data support for subsequent parameter extraction and use. It is the core algorithm for achieving parameter structuring and ensuring parameter accuracy, and unlike conventional table parsing algorithms, it effectively avoids key-value mismatches and parameter confusion.

[0112] Structured PCB parameter basic data refers to a standardized set of PCB parameters formed by key-value matching and structured recombination through parameter key-value association algorithms. The parameter items and their corresponding values ​​are in a one-to-one correspondence and have a standardized format. It serves as the basic data carrier for subsequent parameter retrieval.

[0113] It should be noted that the illogical document data identified by OCR is re-associated with parameter names and specific values ​​by combining visual spatial logic.

[0114] Specifically, a parameter key-value association algorithm is adopted. Based on the table header affiliation and row and column correspondence information recorded in the document global logical architecture data used to constrain the semantic consistency of the text sequence, the parameter names and corresponding values ​​in the corrected text sequence data are matched in a targeted manner. The parameter item is used as the "key" and the corresponding value or identifier is used as the "value" to establish a preliminary mapping relationship. Then, the attribution logic of the key-value pairs is verified according to the spatial association rules of the global logical architecture data, and misaligned or invalid combinations are eliminated. Subsequently, the valid key-value pairs are regularized and restructured according to the industry standard format of PCB parameters to output the basic data of structured PCB parameters.

[0115] Step 105: When a user's PCB parameter retrieval request is received, a vector retrieval model is used to filter noise from the structured PCB parameter base data and the user's PCB parameter retrieval request, and output high-quality document fragment data.

[0116] A user PCB parameter retrieval request refers to a query command submitted by a user to obtain parameter information related to PCB design and manufacturing. It includes the retrieval requirements such as the name, specifications, and attributes of the target parameter.

[0117] Vector retrieval models refer to retrieval models that can map text data into vector form and achieve data matching and filtering through vector similarity calculation. They are used to complete the semantic association between retrieval requests and structured data, as well as noise filtering.

[0118] High-quality document fragment data refers to structured PCB parameter data fragments that have a high semantic matching degree with the user's search request and reliable data association after being filtered out of irrelevant noise data by a vector retrieval model.

[0119] It should be noted that when a user's PCB parameter retrieval request is received, a vector retrieval model is used to map the user's retrieval request and the structured PCB parameter basic data into vector form. Through vector similarity calculation, parameter data fragments that are highly relevant to the request semantics are selected, irrelevant noise data is removed, and high-quality document fragment data is output.

[0120] Furthermore, step 105 may include the following sub-steps:

[0121] S51. A vector retrieval model is used to perform vector similarity matching between the structured PCB parameter basic data and the user's PCB parameter retrieval request to obtain multiple candidate document fragments.

[0122] S52. Perform information density scoring on multiple candidate document fragments to obtain the information density score for each candidate document fragment.

[0123] S53. Compare the information density score of each candidate document fragment with the preset dynamic scoring threshold;

[0124] S54. Select any candidate document fragment data corresponding to an information density score that is greater than or equal to a preset dynamic scoring threshold as high-quality document fragment data.

[0125] It should be noted that during PCB document parsing, OCR recognition noise, repetitive headers and footers, or excessive unstructured descriptions can lead to low-quality content filling the retrieval chunks of the RAG (Retrieval-Augmented Generation System). This solution aims to quantify the effective knowledge value of text chunks, eliminate low-value noise before reordering, and ensure that the model can focus on core data with real physical parameters.

[0126] Furthermore, this invention formalizes the slice filtering process as an adaptive scoring and admission system, the specific definition and process of which are as follows:

[0127] Regarding the definition of Information Density Score (IDS): This invention focuses not only on text length but also on the effective information contained in the text. The IDS is used to control the admission criteria for candidate segments to enter the refined reordering stage. The calculation of IDS is mainly based on the following three core dimensions for comprehensive quantitative evaluation:

[0128] 1) Key parameter weights (W) key ): Identify whether the text contains core design parameter words such as Layer, PTH, Impedance, Stack-up, etc.

[0129] 2) Dimensional density (W) unit : Identify the frequency of occurrence of PCB-specific units (such as um, mil, oz, mm).

[0130] 3) Structured Feature Ratio (P) struct ): Calculates the ratio of numbers, symbols, and ordinary characters in text. Since the core parameters of a PCB are usually in the form of "key-value pairs" or "tables," a high density of numbers usually means higher information value.

[0131] For each document segment C i Its information density score (IDS) i The definition is as follows:

[0132] ;

[0133] in, and These are the word frequencies of keywords and units, respectively. , This is a preset domain importance coefficient; This represents the total length of the fragments, used for normalization. The information entropy of the text is used to penalize a large amount of useless repetitive information parsed from the document—the higher the text repetition, the lower the information entropy. By taking the reciprocal, the weight of high information entropy (i.e., non-repetitive content) is increased in reverse.

[0134] Furthermore, such as Figure 2-3 As shown, the dynamic filtering process consists of three steps:

[0135] 1. Initial screening and scoring: After the vector retrieval recalls the initial Top-K segments, the IDS of each segment is calculated immediately.

[0136] 2. Noise Removal: Set a dynamic threshold to remove fragments where IDS < the dynamic threshold. This effectively filters out a large amount of useless repetitive information in the extracted text.

[0137] 3. Refined Reordering: The filtered high-quality fragments are fed into the Rerank model to obtain high-precision PCB design parameter retrieval data, ensuring that the model resources are centrally processed to contain content with real physical parameters.

[0138] Step 106: Perform fine-grained rearrangement of the high-quality document fragment data to output high-precision PCB design parameter retrieval data.

[0139] High-precision PCB design parameter retrieval data refers to a standardized set of PCB design parameters that has undergone refined rearrangement and is highly relevant to the user's search request, with strong data credibility and accurate parameter matching. It is the final retrieval result that meets the user's parameter query needs.

[0140] It's important to note that the process of fine-grained rearrangement of high-quality document fragments involves directly inputting the fragmented data into the Rerank model. The model combines the semantic features of the user's PCB parameter retrieval request with the parameter content of the fragmented data to calculate the precise match between each fragment and the retrieval request. Based on the match score, the high-quality document fragments are prioritized and their order adjusted to ensure that the parameters most relevant to the user's needs are displayed in the optimal position, ultimately outputting high-precision PCB design parameter retrieval data. By introducing the Rerank model for fine-grained rearrangement, the accuracy of the retrieval results is optimized, effectively solving the problem of missing key information due to coarse result sorting and inaccurate relevance in traditional retrieval methods. This further ensures the accuracy and practicality of the final output PCB design parameter retrieval data.

[0141] Among them, the Reranking Model is a dedicated post-processing model in the Retrieval Enhancement Generation (RAG) architecture used to optimize retrieval results. In this invention, a lightweight semantic re-ranking model is adopted for fine-tuning parameter text adapted to the PCB domain. This model takes high-quality document fragments from user PCB parameter retrieval requests and vector retrieval outputs as input. Through a pre-trained domain-adaptive semantic matching network, it calculates the fine-grained semantic relevance score between each fragment and the retrieval request. At the same time, it combines the professional matching rules of PCB parameters (such as the correspondence between parameter items and values, and the matching degree of industry terms) to generate accurate relevance weights. Based on these weights, the fragments are prioritized and their positions are adjusted. It is the core model carrier for realizing fine-grained re-ranking and optimizing the matching degree of retrieval results.

[0142] It is worth noting that, unlike traditional vector retrieval which relies solely on embedded vector matching and cannot distinguish the effectiveness of content, this invention introduces Information Density Score (IDS) to quantitatively evaluate all candidate segments. It rigorously filters out low-value segments with excessively low effective parameter density and high redundant noise, effectively addressing the core pain point that even if the underlying data is correctly identified, low-information-density segments can still interfere with the Large Language Model (LLM) judgment, leading to content illusion and the omission of key PCB parameters. By using the information density score as the segment admission criterion, deep noise filtering is achieved, ultimately resulting in... We generate high-quality document fragments with concise content and rich core parameters. Based on this, we conduct refined reordering of the high-quality document fragments. Through a reordering model combined with semantic features and parameter association rules in the PCB field, we perform fine-grained relevance scoring and order optimization on the high-information-density retention fragments. On the basis of high-density effective content, we further calibrate the search matching priority, forming a progressive optimization link of coarse-grained vector screening + information density admission filtering + fine-grained semantic reordering. This reduces the interference of invalid information layer by layer, focuses on core parameter content, and comprehensively improves the accuracy and content quality of search results.

[0143] For comparison of technical effectiveness, existing technologies can be referenced. In the field of printed circuit board (PCB) design and manufacturing, engineers need to accurately fill in hundreds of design parameters based on a vast amount of industry specifications, internal company standards, and board-level documentation. Traditional search-enhanced generation (RAG) technology often performs poorly in handling such tasks due to inaccurate input sources and overly lengthy search content. To address these issues, this patent proposes a PCB parameter retrieval technology based on multimodal semantic verification and information density filtering, enabling accurate searching of PCB design parameters.

[0144] Specifically, traditional Retrieval Enhanced Generation (RAG) technology performs poorly when handling such high-precision industrial tasks. On the one hand, general OCR (Optical Character Recognition) technology is prone to physical and logical-level recognition biases when dealing with professional micro-annotations, lead relationships, and complex layouts in the PCB field (such as misidentifying the key copper thickness parameter 1oz as 10z, resulting in errors in the source data of the retrieval system and seriously affecting the accuracy of the final parameter extraction). On the other hand, because PCB documents are filled with a large number of repetitive headers and footers and unstructured redundant descriptions, the RAG system is easily interfered with by low-value fragments during the retrieval stage, making it difficult to accurately focus on core parameter items with real physical meaning from massive background information.

[0145] Furthermore, mainstream solutions primarily rely on the traditional RAG architecture and general embedding models. These models segment documents into fragments and store them in a quantized format. When a user asks a question, the system retrieves the most relevant fragments using vector similarity and integrates them with a large model (LLM) to generate the answer. At the semantic understanding level, existing solutions typically utilize embedding models combined with supervised contrastive learning strategies. This enhances the model's recall accuracy for specialized terms and their context by bringing positive sample pairs closer together and pushing away difficult negative samples in the vector space.

[0146] Based on the above, the main shortcomings of the existing technology are: (1) Limited accuracy of industrial document parsing and lack of logical verification. When traditional OCR technology processes professional micro-annotations, Gerber document annotations and high-density SPEC tables in the PCB field, due to the lack of domain knowledge constraints, it is very easy to produce key character misidentification (such as mistaking 1oz for 10z) or page layout logical misalignment (such as table row and column misalignment). Such source document recognition deviation will cause the retrieval system to receive incorrect underlying data, making the subsequent parameter extraction lose its accuracy basis. (2) Severe noise interference in retrieval fragments. The existing RAG architecture mainly relies on general vector similarity for matching, and has not established a quantitative filtering mechanism for low-value information such as repeated headers and footers, unstructured descriptions and legal disclaimers that frequently appear in PCB documents. As a result, the retrieval results (Top-K) are often filled with a large number of noisy fragments that lack core physical parameters, which not only dilutes the effective information, but also seriously interferes with the reasoning efficiency and final accuracy of large models when processing complex parameter tracing tasks.

[0147] To address the limitations in document parsing accuracy and redundant noise interference in PCB design parameter retrieval, this invention proposes a PCB parameter retrieval method based on multimodal semantic verification and information density filtering. By collaboratively correcting biases using OCR and a visual large model (VLM), the quality of source data is improved. Furthermore, information density scores (IDS) are used to quantify and eliminate low-value noise fragments. This aims to enhance the accuracy and efficiency of the large model in extracting design parameters from a multi-source heterogeneous PCB knowledge base, achieving high-fidelity parameter traceability.

[0148] Specifically, this invention primarily addresses typical application scenarios in the design and manufacturing of printed circuit boards (PCBs), where engineers perform parameter retrieval and information tracing in complex, multi-source document environments. The core challenges lie in two main aspects: First, traditional document parsing technologies (such as general OCR) are insufficiently accurate in extracting PCB-related terminology, complex table layouts, and minute annotations, resulting in noisy source data entering the retrieval system and severely impacting the accuracy of final parameter extraction. Second, existing Retrieval Enhancement Generation (RAG) technologies, when processing PCB documents containing a large amount of redundant and unstructured descriptions, easily recall low-quality noisy fragments, interfering with the model's accurate extraction of core design parameters. Therefore, this patent's technical description is divided into two parts: one part is a high-precision text extraction and semantic verification technology based on the collaboration of OCR and Visual Large Model (VLM), aiming to improve the accuracy of the input through multimodal cross-validation; the other part is a source document noise filtering technology based on Information Density Score (IDS), which dynamically eliminates low-value noisy data during the retrieval stage by quantifying the physical parameter value and information entropy of text fragments. By organically combining these two parts, a PCB parameter retrieval technology based on multimodal semantic verification and information density filtering is formed, enabling efficient and accurate tracing of complex PCB design parameters.

[0149] Compared with existing parameter retrieval techniques based on general embedding models and traditional RAG architectures, this invention has the following advantages:

[0150] 1. High-precision physical logic verification and source correction capabilities: Through the collaboration of OCR and Visual Model (VLM), and by utilizing the spatial perception and domain knowledge constraints of VLM, semantic verification and physical logic correction (such as dimensional error correction and table alignment) of the original extraction results are achieved, effectively solving the problem of parameter extraction distortion caused by input end recognition deviation in traditional technologies.

[0151] 2. Highly efficient document noise resistance and refined retrieval performance: By introducing a quantitative filtering mechanism based on Information Density Score (IDS), and through a comprehensive evaluation of key parameter weights and information entropy, the system can accurately remove redundant and unstructured noise in PCB documents, ensuring that the system focuses on core data with real physical meaning, and significantly improving the retrieval fidelity in complex contexts.

[0152] In this embodiment of the invention, a PCB parameter retrieval method based on multimodal semantic verification and information density filtering is provided. This method acquires multi-source heterogeneous PCB document data and performs multimodal layout analysis on the data, outputting modal slice data. A pre-trained OCR model and a pre-trained VLM model are used to perform parallel perceptual extraction on the modal slice data, outputting OCR text sequence data and document global logical architecture data for constraining the semantic consistency of the text sequence. A MAP model is used to perform physical logical correction on the OCR text sequence data based on a prior knowledge base in the PCB domain and the document global logical architecture data for constraining the semantic consistency of the text sequence, outputting corrected text sequence data. A parameter key-value association algorithm is used to perform structured data synthesis on the corrected text sequence data and the document global logical architecture data for constraining the semantic consistency of the text sequence, outputting structured PCB parameter basic data. When a user's PCB parameter retrieval request is received, a vector retrieval model is used to retrieve the structured PCB parameter basic data and the user's request. The PCB parameter retrieval request undergoes noise filtering to output high-quality document fragment data. This high-quality document fragment data is then finely rearranged to output high-precision PCB design parameter retrieval data. Based on this approach, this invention leverages multimodal layout analysis to provide well-structured slice data for subsequent data processing. It extracts both text characters and document spatial logic information through parallel OCR and VLM extraction. Then, using a MAP model combined with domain prior knowledge and a global logical architecture, it completes physical logic correction, fundamentally correcting character misidentification and layout misalignment issues caused by a lack of domain constraints in traditional methods. This eliminates erroneous underlying data resulting from source document recognition bias, laying a solid data foundation for accurate parameter extraction. Combined with a parameter key-value association algorithm, it achieves structured data synthesis, further ensuring data standardization and effectiveness. Subsequently, a vector retrieval model filters noise, and by quantifying the physical parameter value and information entropy of the text fragments, low-value noise data is dynamically removed during the retrieval stage. This effectively solves the problem of low effective parameter density in the fragments, which leads to problems with large model (LLM). The problem of language models being misled by noise, resulting in illusions or omissions of key parameters is addressed through refined rearrangement and optimization, which effectively ensures the accuracy and reliability of PCB design parameter extraction results.

[0153] Please see Figure 4 , Figure 4 This is a structural block diagram of a PCB parameter retrieval system based on multimodal semantic verification and information density filtering, provided in Embodiment 2 of the present invention.

[0154] This invention provides a PCB parameter retrieval system based on multimodal semantic verification and information density filtering, comprising:

[0155] The acquisition module 401 is used to acquire PCB multi-source heterogeneous document data, perform multimodal layout analysis on the PCB multi-source heterogeneous document data, and output modal slice data.

[0156] The extraction module 402 is used to perform parallel perceptual extraction of modal slice data using a pre-trained OCR model and a pre-trained VLM model, and output OCR text sequence data and document global logical architecture data for constraining the semantic consistency of the text sequence.

[0157] The correction module 403 is used to perform physical and logical correction on the OCR text sequence data using the MAP model based on the prior knowledge base of the PCB domain and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and output the corrected text sequence data.

[0158] The synthesis module 404 is used to perform structured data synthesis on the corrected text sequence data and the document global logical architecture data used to constrain the semantic consistency of the text sequence using a parameter key-value association algorithm, and output the basic data of structured PCB parameters.

[0159] The filtering module 405 is used to filter noise from the structured PCB parameter basic data and the user PCB parameter retrieval request using a vector retrieval model when a user PCB parameter retrieval request is received, and output high-quality document fragment data.

[0160] The rearrangement module 406 is used to perform fine-grained rearrangement of high-quality document fragment data and output high-precision PCB design parameter retrieval data.

[0161] Optionally, the acquisition module is specifically used for:

[0162] Region recognition is performed on multi-source heterogeneous PCB document data to obtain text areas, parameter table areas, and Gerber visual annotation areas;

[0163] Based on text areas, parameter table areas, and Gerber visual annotation areas, the multi-source heterogeneous PCB document data is partitioned to obtain the corresponding region data for the text areas, parameter table areas, and Gerber visual annotation areas.

[0164] By integrating the region data corresponding to the text area, parameter table area, and Gerber visual annotation area, modal slice data is obtained.

[0165] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0166] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the PCB parameter retrieval method based on multimodal semantic verification and information density filtering as described in the above embodiments.

[0167] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the PCB parameter retrieval method based on multimodal semantic verification and information density filtering as described in the above embodiments.

[0168] This invention also provides a computer program product, including a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer performs the steps of the PCB parameter retrieval method based on multimodal semantic verification and information density filtering as described in the above embodiments.

[0169] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0170] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0171] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0172] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0173] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A PCB parameter retrieval method based on multimodal semantic verification and information density filtering, characterized in that, include: Acquire PCB multi-source heterogeneous document data, perform multimodal layout analysis on the PCB multi-source heterogeneous document data, and output modal slice data; The modal slice data is extracted in parallel using a pre-trained OCR model and a pre-trained VLM model, and the output is OCR text sequence data and document global logical architecture data used to constrain the semantic consistency of the text sequence. The MAP model is used to perform physical and logical correction on the OCR text sequence data based on the prior knowledge base of the PCB domain and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and the corrected text sequence data is output. A parameter key-value association algorithm is used to perform structured data synthesis on the corrected text sequence data and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and the structured PCB parameter basic data is output. When a user PCB parameter retrieval request is received, a vector retrieval model is used to filter noise from the structured PCB parameter base data and the user PCB parameter retrieval request, and output high-quality document fragment data. The high-quality document fragment data is finely rearranged to output high-precision PCB design parameter retrieval data.

2. The PCB parameter retrieval method based on multimodal semantic verification and information density filtering according to claim 1, characterized in that, The process of performing multimodal layout analysis on the PCB multi-source heterogeneous document data and outputting modal slice data includes: Region recognition is performed on the PCB multi-source heterogeneous document data to obtain text area, parameter table area and Gerber visual annotation area; Based on the text area, the parameter table area, and the Gerber visual annotation area, the PCB multi-source heterogeneous document data is divided into data segments to obtain the corresponding region data for the text area, parameter table area, and Gerber visual annotation area. By integrating the region data corresponding to the text area, the parameter table area, and the Gerber visual annotation area, modal slice data is obtained.

3. The PCB parameter retrieval method based on multimodal semantic verification and information density filtering according to claim 1, characterized in that, The modal slice data is extracted in parallel using a pre-trained OCR model and a pre-trained VLM model, outputting OCR text sequence data and document global logical architecture data for constraining the semantic consistency of the text sequence, including: The pre-trained OCR model is used to perform character-level scanning and extraction on the modal slice data to obtain the original OCR text sequence data; The original OCR text sequence data is input into the pre-trained VLM model for review; If the review fails, the pre-trained VLM model is used to correct the original OCR text sequence data to obtain OCR text sequence data. If the review is approved, the original OCR text sequence data will be used as the OCR text sequence data. The pre-trained VLM model is used to perform spatial logic perception processing on the modal slice data to obtain the table header membership relationship and the leader pointing information; By integrating the table header membership relationship and the leader pointing information, a document global logical architecture data is obtained to constrain the semantic consistency of the text sequence.

4. The PCB parameter retrieval method based on multimodal semantic verification and information density filtering according to claim 1, characterized in that, The MAP model is used to perform physical logic correction on the OCR text sequence data based on the prior knowledge base of the PCB domain and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and outputs the corrected text sequence data, including: The MAP model is used to solve the optimal text sequence problem on the OCR text sequence data based on the document global logical architecture data used to constrain the semantic consistency of the text sequence and the PCB domain prior knowledge base, so as to obtain the solved text sequence. Based on the prior knowledge base in the PCB field, character error correction is performed on the solved text sequence to obtain the error-corrected text sequence. The document global logical architecture data used to constrain the semantic consistency of the text sequence is used to perform table row and column alignment on the corrected text sequence to obtain the corrected text sequence data.

5. The PCB parameter retrieval method based on multimodal semantic verification and information density filtering according to claim 1, characterized in that, The vector retrieval model is used to filter noise from the structured PCB parameter base data and the user's PCB parameter retrieval request, outputting high-quality document fragment data, including: The vector retrieval model is used to perform vector similarity matching between the structured PCB parameter base data and the user's PCB parameter retrieval request, resulting in multiple candidate document fragments. Information density scores are performed on multiple candidate document fragments to obtain an information density score for each candidate document fragment. The information density score of each candidate document fragment is compared with a preset dynamic scoring threshold. Any candidate document fragment data corresponding to an information density score greater than or equal to the preset dynamic scoring threshold is used as high-quality document fragment data.

6. A PCB parameter retrieval system based on multimodal semantic verification and information density filtering, characterized in that, include: The acquisition module is used to acquire PCB multi-source heterogeneous document data, perform multimodal layout analysis on the PCB multi-source heterogeneous document data, and output modal slice data. The extraction module is used to perform parallel perceptual extraction of the modal slice data using a pre-trained OCR model and a pre-trained VLM model, and output OCR text sequence data and document global logical architecture data for constraining the semantic consistency of the text sequence. The correction module is used to perform physical and logical correction on the OCR text sequence data using the MAP model based on the PCB domain prior knowledge base and the document global logical architecture data used to constrain the semantic consistency of the text sequence, and output the corrected text sequence data. The synthesis module is used to perform structured data synthesis on the corrected text sequence data and the document global logical architecture data used to constrain the semantic consistency of the text sequence using a parameter key-value association algorithm, and output structured PCB parameter basic data. The filtering module is used to filter noise from the structured PCB parameter base data and the user PCB parameter retrieval request using a vector retrieval model when a user PCB parameter retrieval request is received, and output high-quality document fragment data. The rearrangement module is used to perform fine-grained rearrangement of the high-quality document fragment data and output high-precision PCB design parameter retrieval data.

7. The PCB parameter retrieval system based on multimodal semantic verification and information density filtering according to claim 6, characterized in that, The acquisition module is specifically used for: Region recognition is performed on the PCB multi-source heterogeneous document data to obtain text area, parameter table area and Gerber visual annotation area; Based on the text area, the parameter table area, and the Gerber visual annotation area, the PCB multi-source heterogeneous document data is divided into data segments to obtain the corresponding region data for the text area, parameter table area, and Gerber visual annotation area. By integrating the region data corresponding to the text area, the parameter table area, and the Gerber visual annotation area, modal slice data is obtained.

8. An electronic device, characterized in that, The device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the PCB parameter retrieval method based on multimodal semantic verification and information density filtering as described in any one of claims 1-5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the PCB parameter retrieval method based on multimodal semantic verification and information density filtering as described in any one of claims 1-5.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the steps of the PCB parameter retrieval method based on multimodal semantic verification and information density filtering as described in any one of claims 1-5.

Images