A qualified certificate intelligent analysis and management system based on evidence cooperation

Abstract

This invention discloses an intelligent analysis and management system for certificates of conformity based on evidence collaboration, comprising a document analysis module, an evidence collaboration module, a trusted evidence storage module, a business linkage module, and a visualization governance module. The document analysis module performs structured analysis of certificate images and generates field content and confidence levels. The evidence collaboration module establishes and verifies the correspondence between certificates of conformity and physical products based on RFID / NFC tags, production chain data, and environmental chain data. The trusted evidence storage module uses blockchain technology to write operation records into a trusted ledger, achieving end-to-end evidence storage. The business linkage module automatically triggers business processes such as warehousing, release, and review based on consistency scores and on-chain compliance status, and records business behaviors. The visualization governance module constructs an event timeline and entity relationship diagram, providing comprehensive data presentation and traceability. This invention achieves integrated management of certificate of conformity information, improving the accuracy, reliability, and traceability of certificate of conformity management.

Description

Technical Field

[0001] This invention relates to the field of industrial quality management and information processing technology, specifically to an intelligent analysis and management system for certificates of conformity based on evidence-material collaboration. Background Technology

[0002] In the process of industrial manufacturing and product quality management, the certificate of conformity serves as an important document for recording product inspection results, production information, and quality status. Its accuracy, uniqueness, and traceability are crucial for ensuring product quality and guaranteeing supply chain stability.

[0003] In existing technologies, certificates of conformity are typically in paper or semi-digital form. Enterprises often rely on traditional methods such as manual data entry, manual verification, and distributed storage for management. However, due to the complex shooting conditions in industrial settings and the diverse formats and structures of certificates, existing document recognition technologies are prone to inaccurate identification, incomplete structural parsing, missing or incorrect fields, and other problems when processing such certificates. This makes it difficult to meet the requirements of quality management for accuracy, completeness, and stability. Meanwhile, products generate a large amount of physical information and status records from equipment, tags, or IoT devices during production, warehousing, distribution, and use. However, this information is usually independent of the certificate content and lacks effective correlation. Existing systems generally cannot continuously, dynamically, and reliably verify the correspondence between the certificate and the product itself, leading to common problems such as "certificate and product disconnect," "certificate accompanying but not accompanying product," and "mixed use of certificates and products." Especially in supply chain scenarios involving multiple stages and multiple stakeholders, paper or image-based certificates are more easily replaced, mismatched, or unable to be verified for consistency with the product, making quality traceability and liability determination difficult.

[0004] Therefore, there is an urgent need for a comprehensive system that can simultaneously address certificate identification and physical product verification, enabling document parsing results and entity identification data to be integrated and processed within the same framework. This system will allow for the accurate extraction of certificate information and the reliable acquisition of physical product information, and establish a credible relationship between the two, thereby solving long-standing problems in existing technologies such as inaccurate identification, discrepancies between certificates and physical products, and difficulties in traceability. Summary of the Invention

[0005] The objective of this invention is to propose an intelligent analysis and management system for certificates of conformity based on evidence collaboration, thereby solving the problems mentioned in the background art.

[0006] The system comprises five modules: document parsing, physical evidence collaboration, trusted evidence storage, business linkage, and visual governance.

[0007] The document parsing module uses a multi-model table classification algorithm to determine the table type of the input certificate image; based on the determination result, it selects either a wired table structure parsing algorithm based on U-Net or a wireless table structure parsing algorithm based on a two-stage Transformer architecture to generate a table structure; and combines a multi-OCR fusion recognition engine with template verification to output the structured field content and field confidence of the certificate image. The physical evidence collaboration module reads RFID / NFC / QR code identifiers and collects multi-source sensor data, and performs cross-source consistency verification between the structured fields and the product's physical production chain, environmental chain, and signal feature sequences to form a consistency score. The trusted evidence storage module adopts a hierarchical hash evidence structure and a credit-weighted consensus mechanism to write field content, evidence verification results and business behavior into a trusted ledger using visualized smart contract rules. The business linkage module automatically triggers or blocks enterprise business processes based on consistency scores and on-chain compliance records. The visualization governance module presents the document parsing path, physical evidence verification link, on-chain evidence structure and business actions in the form of an event timeline and entity relationship diagram, forming a governance interface based on the evidence chain.

[0008] Furthermore, the above system also includes the following: In complex industrial environments, the document parsing module recovers text regions, perspective distortion, lighting imbalance, and local blur through image preprocessing. It accurately determines table types using a three-model joint mechanism (YOLO, PaddleClas, and Qanything), then follows the table type to either a wired table's grid reconstruction path or a LORE two-stage Transformer semantic structure inference path for wireless tables. Through U-Net probabilistic graphs and graph theory connectivity strategies, or semantic structure inference via local-global attention fusion, the system ultimately obtains accurate cell rectangle structures and text regions. Through parallel recognition by multiple OCR engines, confidence fusion, and semantic verification, the system obtains field values ​​and performs format, semantic, and contextual consistency checks. A template matching mechanism further confirms whether the certificate format meets enterprise or industry requirements and provides semantic anchors for subsequent verification and evidence preservation. Finally, it outputs structured data, field confidence chains, template matching results, model versions, and intermediate processing evidence, which constitute the system's initial digital certificate form and also provide verification basis for downstream modules.

[0009] The physical evidence collaboration module precisely binds certificates to physical products. The UID serves as a unified physical evidence identifier across systems, devices, and enterprises. It is used to read production chain records from MES, PLC, and industrial control data streams, as well as environmental data such as temperature, humidity, and vibration from the warehouse network. By comparing the UID data obtained through multi-mode IoT readers with fields parsed from the document, the system determines whether the certificate truly corresponds to the physical object. In this process, the system can detect OCR recognition errors, as well as risks such as mislabeling, label replacement, product substitution, and environmental non-compliance, thus providing a "physical consistency score" and packaging the consistency evidence into hash groups to form a digital image of the real world.

[0010] The trusted evidence storage module writes all evidence from the document parsing module and the physical evidence collaboration module into the industrial lightweight consortium blockchain in a way that is verifiable across organizations, devices, and platforms. Before being written to the blockchain, the evidence is organized into three logical levels: identity, process, and compliance, forming an immutable evidence tree through a multi-hash structure and referencing methods. The on-chain content includes not only the hashes of field values ​​but also field confidence levels, model versions, IoT physical layer characteristics, production chain summaries, warehousing chain summaries, consistency conclusions, and risk markers, thereby achieving multi-source proof between certificates of conformity, physical objects, environments, and behaviors. The consortium blockchain uses credit-weighted consensus and visual smart contracts, using enterprise credit, equipment collection quality, and historical behavior credibility as consensus weights. The smart contracts not only verify the evidence format but also execute causal rules across events, such as verifying whether the manufacturing date is later than the process completion time, whether the warehousing environment meets the certificate of conformity declaration conditions, and whether business events occur in sequence. Every record on the blockchain becomes verifiable evidence for future audits and judicial disputes, giving the entire system legal validity.

[0011] Once the on-chain write is complete, the business linkage module begins translating the complex evidence into actual enterprise actions. Through semantic mapping, the module maps structured evidence to ERP material master data, MES process information, and WMS warehouse entities, converting consistency scores and trusted evidence markers into business risk weights. Based on this, the system automatically determines business actions: if the evidence is consistent, on-chain compliant, and the confidence level is stable, actions such as warehousing, shelving, and release are automatically executed; if minor anomalies exist, sampling inspection or enhanced review is implemented; if cross-source contradictions occur, the business is automatically frozen and an alarm is triggered. This module can also receive verification events from upstream enterprises in the trusted ledger, automating supply chain collaboration, such as automatically generating downstream receiving tasks after upstream verification. The execution logs of all business actions are written back to the blockchain to form behavioral evidence, ensuring that business execution is auditable and non-repudiable.

[0012] The visualization governance module re-integrates all identification, verification, evidence storage, and business actions into a governance visualization graph. The system displays the entire process of a certificate of conformity from photographing, parsing, physical evidence matching, evidence storage, to business execution in an event timeline format, clearly presenting the evidence and causal relationships at each stage. It uses a graph-based approach to display the reference relationships between the certificate itself, physical evidence UID, production workstation, warehouse node, on-chain hash, and business actions, making risk clusters, abnormal links, process deviations, and warehouse anomalies visually apparent. The module also provides governance intervention capabilities, allowing managers to directly adjust thresholds, strategies, or process nodes through the governance interface and write all adjustment records on the blockchain to form a governance evolution chain. Visualized governance not only displays the current state but also combines document recognition confidence trends, IoT fluctuation trends, and warehouse environment sequences to predict trends, thereby identifying potential risks in advance.

[0013] This invention provides an intelligent analysis and management system for certificates of conformity based on evidence collaboration. Compared with existing technologies, it has the following advantages: This invention first introduces a multi-model fusion structured parsing mechanism at the document parsing level. Through table type classification, template matching, LORE wireless table structure inference, joint recognition by multiple OCR engines, and semantic consistency verification, the system can achieve high-precision and robust parsing of certificate fields even under complex industrial shooting conditions and with numerous table styles. Compared with traditional OCR or single structured methods, it can output structural semantics, field confidence chains, model evidence, and template matching results, making the document content easier for subsequent verification, evidence preservation, and business modules to directly utilize.

[0014] Secondly, this invention constructs a physical evidence collaboration module to perform real-time matching and consistency verification between the certificate of conformity parsing results and physical source information such as RFID / NFC tags, production equipment data, and warehousing environment characteristics, achieving a natural extension of evidence from the document space to the physical space. The system not only verifies whether the field content is consistent with the product's production chain, warehousing chain, and equipment chain, but also constructs physical-level digital evidence through signal feature sequences, environmental time series, and link summaries, fundamentally solving long-standing problems in traditional systems such as "inconsistent evidence and physical goods," "evidence accompanying but not accompanying the goods," and "substitutable evidence."

[0015] Regarding data trustworthiness, this invention utilizes a trusted evidence storage module to hierarchically hash document identification evidence, physical evidence verification evidence, and business behavior evidence according to identity, process, and compliance layers, and then writes this information into a lightweight industrial consortium blockchain. The system employs credit-weighted consensus and visual smart contracts to perform rule-based verification on cross-modal evidence, ensuring that every identification, matching, and process action possesses immutable, verifiable, and traceable on-chain attributes. Compared to existing methods of storing single documents or recording partial events on the blockchain, this invention constructs a complete evidence chain encompassing certificates of conformity, physical objects, production records, warehousing environment, and business behavior.

[0016] In terms of business collaboration, this invention realizes an automatic business triggering mechanism for document parsing results and physical evidence verification conclusions through a business linkage module, which can interact bidirectionally with systems such as ERP, MES, and WMS. The system automatically generates business actions such as warehousing, freezing, release, and review based on field confidence, consistency score, and on-chain compliance results, and writes the receipts into a trusted ledger, so that the business process is highly integrated with the identification and verification results. This mechanism effectively eliminates the "data silo" problems in traditional systems, such as document recognition not being able to drive business, physical evidence data not being able to flow back to business, and each system being independent.

[0017] At the governance level, this invention introduces a visualization governance module based on the chain of evidence, which visualizes the document parsing process, physical evidence reading link, on-chain evidence storage structure, and business decision-making logic in a holistic manner according to time, causality, and entity relationships. This enables managers to uniformly monitor the entire lifecycle of the certificate of conformity from generation to use. The system also provides visualized policy control, anomaly cluster identification, risk trend prediction, and governance evolution chain, transforming quality management from static review to a dynamic, interventionist, and evolvable governance system. This capability is currently unattainable in existing certificate of conformity management technologies.

[0018] In summary, this invention achieves a smarter, more reliable, more streamlined, and more auditable upgrade of certificate management through the deep integration of five aspects: cross-modal data analysis, real-time physical collaboration, construction of a trusted chain of evidence, automated business linkage, and overall governance visualization. This significantly improves the efficiency, accuracy, security, and regulatory value of industrial quality management, and effectively solves long-standing problems in existing technologies such as inaccurate identification, discrepancies between certificates and physical objects, scattered evidence storage, fragmented business processes, and lack of governance visibility. Attached Figure Description

[0019] Figure 1 This is a structural diagram of a certificate of conformity intelligent analysis and management system based on evidence collaboration. Detailed Implementation

[0020] 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, and 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.

[0021] In one or more embodiments, the present invention provides as follows Figure 1 The illustrated intelligent certificate analysis and management system based on evidence collaboration includes the following structure: Document parsing module: Determines the table type of the input certificate image using a multi-model table classification algorithm; selects either a wired table structure parsing algorithm based on U-Net or a wireless table structure parsing algorithm based on a two-stage Transformer architecture based on LORE based on the judgment result, and then generates the table structure; combines a multi-OCR fusion recognition engine with template verification to output the structured field content and field confidence of the certificate image. The physical evidence collaboration module reads RFID / NFC / QR code identifiers and collects multi-source sensor data. It then performs cross-source consistency verification between the structured fields and the product's physical production chain, environmental chain, and signal feature sequences to form a consistency score. Trusted Evidence Storage Module: Employs a layered hash evidence structure and a credit-weighted consensus mechanism to write field content, evidence verification results, and business behaviors into a trusted ledger using visualized smart contract rules; Business linkage module: Automatically triggers or blocks enterprise business processes based on consistency scores and on-chain compliance records; Visualized governance module: Presents document parsing paths, physical evidence verification links, on-chain evidence structures, and business actions in the form of event timelines and entity relationship diagrams, forming a governance interface based on the evidence chain.

[0022] Furthermore, the above system also includes the following: In this embodiment, the document parsing module: Image reception and preprocessing stage: Receiving Base64, file, and OpenCV / PIL format images, first performing unified format conversion (Base64 decoded to binary stream and converted to OpenCV BGR format, PIL image converted to RGB format, alpha channel removed and unified to 8-bit depth); reading EXIF ​​orientation information to perform preliminary rotation correction. If the image tilt angle after correction is >1°, further geometric correction is performed: edges are extracted using Canny edge detection (low threshold 50, high threshold 150), combined with Hough line transform (ρ precision 1 pixel, θ precision 1°, voting threshold 100) to obtain a set of lines, using K-means clustering (angle difference <3° is considered the same class) to estimate the main orientation of the page, and then using RANSAC fitting of a group of lines (at least 20 lines, interior point threshold 2°) as a reference to correct small tilt angles <5°; subsequently, constructing the image binary boundary contour (Otsu Adaptive thresholding is used to calculate the convex hull and filter corner points (removing redundant corner points with a spacing of <10 pixels). For images with perspective distortion (trapezoidal distortion determination: the difference in length between the top and bottom edges >10%), perspective transformation is performed and mapped to an orthographic projection of A4 standard size (210×297mm, 300DPI corresponds to 2480×3508 pixels). The perspective matrix (4×4 array format) is saved so that the recognized coordinates can be back-projected onto the original image later. The image enhancement process first determines the image scene: if the average brightness is <50, it is considered low light, and multi-scale Retinex (scale 15 / 80 / 250) enhancement is used; if the local contrast is insufficient, CLAHE (8×8 block size, contrast limit 2.0) is performed to stretch the local contrast; if the average gradient magnitude of the image is <20, it is considered slightly blurred, and character edges are recovered by blind deconvolution (10 iterations, regularization parameter 0.01). After enhancement, the character edge gradient is verified (if the loss is >10%, it is reverted to the original image); if the image has no valid straight lines / corners, geometric correction is skipped and the reason is noted in the log. All preprocessing steps record a structured transformation log (including image MD5 identifier, rotation angle (retaining 1 decimal place), scaling ratio, perspective matrix, enhancement algorithm and parameters), which is written into the evidence package along with the preprocessed image.

[0023] Table type determination stage: Three parallel models are used for collaborative determination, namely a lightweight table detection model based on the YOLO series, a high-precision table classification model based on PaddleClas, and a special table classification model based on Qanything. The images need to be adapted to different input scales. The YOLO model uses a 640×640 scale, while the PaddleClas and Qanything models use a 224×224 scale. The YOLO detection model outputs class probability vectors (wired / wireless) and bounding box confidence scores, while the PaddleClas and Qanything classification models only output class probability vectors. Temperature scaling is applied to the class probability vectors of the three models to calibrate their probability distributions (the temperature parameter is automatically selected by minimizing the negative log-likelihood loss on the validation set). Simultaneously, a localization consistency score is calculated for the YOLO model (based on the average IOU between its predicted bounding box and the preset template prior bounding box, which is an anchor box generated according to a typical table size). This score is then subjected to "mean-variance normalization" (based on the IOU distribution of all samples in the validation set, ensuring that its mean and variance are fully aligned with the distribution characteristics of the YOLO class probabilities, guaranteeing comparable contributions). The results were then integrated using a "layered weighted fusion" strategy: The first layer first fused the normalized YOLO class probabilities and the localization consistency scores with a weight of 3:1 (based on the impact test of the localization error on the judgment result on the validation set; when the localization error increases by 5%, the misclassification rate increases by 1 / 3 of the class probability error, hence this ratio was determined) to obtain the YOLO comprehensive score; The second layer then fused this score with the class probabilities of PaddleClas and Qanything with a weighted fusion. The initial weights were set according to the F1 values ​​of the three models on the validation set in a ratio of 4:3:3, and were dynamically adjusted according to the misclassification rate within a 7-day sliding window during online operation (for every 1% increase in the misclassification rate, the corresponding model weight decreased by 0.05). After fusion, if the score is higher than the positive class threshold (determined by taking the maximum value of the Youden index from the ROC curve of the validation set, corresponding to a wired table accuracy of 92.3%), it is judged as a wired table; if it is lower than the negative class threshold (determined by maximizing the wireless table recall rate of the validation set, with a value 0.15 lower than the positive class threshold, corresponding to a wireless table recall rate of 91.8%, and the interval is determined based on the overlap test of the two classes of sample distributions), it is judged as a wireless table; if the score is in the uncertain interval between the two thresholds, the sample automatically enters the manual review queue, and the review results are synchronously fed back to the active learning module. The trigger condition is set to "the number of queue samples accumulates to 500" (determined based on the proportion of edge scene samples in the validation set and the incremental training convergence efficiency; 500 samples can cover more than 95% of edge scenes). If it does not reach 500 samples in more than 30 days, it will be forcibly triggered to ensure continuous model iteration.

[0024] Path to parsing wired tables: After identifying a wired table, pixel-level table line detection is first performed, following this process: The original image is preprocessed with denoising (Gaussian filtering, kernel size 3×3) and distortion correction (perspective transformation based on camera intrinsics), then scaled to 1024×1024. For tables spanning multiple sizes (original image side length > 1024 pixels), slicing is used, with a 20-pixel overlap area reserved at the slice edges. When merging slices, duplicates are removed based on the distance between line segment endpoints (<3 pixels are considered the same line) and direction consistency (angle < 5°), retaining line segments with higher confidence to avoid broken or repeated table lines. The preprocessed image is input into a U-Net-style segmentation network. The network uses RGB mean-variance or ImageNet standard normalization to adapt to the pre-trained weights, outputting a two-channel table line probability map (horizontal and vertical lines). The segmentation results are sequentially subjected to Otsu adaptive thresholding (dynamically determining the threshold based on the grayscale distribution of the probability map, retaining more than 95% of the true table lines), morphological closure (using 3×3 rectangular structural elements to fill gaps where the line width is less than 2 pixels), and optimized Zhang-Suen thinning (retaining the core pixels of intersecting lines while iteratively deleting boundary pixels to avoid gaps at intersections). The probability map is then converted into a binary line graph with a line width of 1 pixel, and the initial endpoints of each line segment are extracted. Considering that the initial endpoints may have a 1-2 pixel offset, they are fed into Cycle CenterNet as candidate keypoints. The precise coordinates and offsets of the endpoints are regressed using center regression, while the direction vector of the original line segment is used to constrain the consistency of the regression, ensuring that the regressed endpoints match the line segment direction.A topology graph is constructed based on the corrected endpoints and line segments: nodes are endpoints, and edges are candidate connections "in the same row / column and with endpoint distance < 8 pixels". The edge attributes include the length and orientation angle of the corresponding line segment. Cells are then constructed using graph theory methods. A normalized cost function is designed to score candidate edges (the weight ratio of the three indicators is 3:2:5, all normalized to [0,1]). Specifically, the scoring functions are: orientation consistency score (0 points for line segment angle < 10°, 1 point for > 30°, with linear interpolation in between), endpoint distance penalty (score increases linearly when pixel distance > 5, 0 points for distance within 5 pixels), and local thermal response cost (based on Cycle). The endpoint confidence normalization result output by CenterNet (normalized to [0,1]) sets the cost to 0 when the heat value is >0.8, decreases linearly between 0.5 and 0.8, and doubles when it is below 0.5. Based on this cost, a heuristic minimum cycle search (prioritizing rectangular closed loops with cost <0.3) is used to generate candidate cells. Then, effective cells are selected by IOU (candidate cells are removed if IOU with the segmentation network table line area is <0.6) and shape compactness (rectangularity = actual area of ​​cell / area of ​​minimum bounding rectangle, removed if below 0.8). For the broken line scenario, line segment extension is performed within the cost threshold <0.25: the maximum extension length is set to 10 pixels, the extension direction is determined by fitting the direction vectors of the three adjacent line segments in the same row / column, and the consistency between the extension path and the table line probability map is calculated at the pixel level (probability >0.6 is considered a valid extension) to avoid generating false connectivity.

[0025] Wireless table parsing path: After determining it to be a wireless table, the table structure is inferred based on text layout and blanking rate. A LORE two-stage Transformer architecture is used, with the specific process as follows: First, OCR text block detection is performed on the page image (using the EAST text detector to locate text regions, outputting the coordinates, font size, and character content of each text block). Then, it is divided into overlapping local windows of 256×256 pixels (overlap ratio set to 20% to avoid text block fragmentation across windows). Each window inputs a local encoder built on a lightweight Transformer, arranging the text blocks from top left to bottom right within the window. The relative position of the text blocks (pixel offset relative to the top left corner of the window) and font size (in pixels) are encoded. The first stage uses the local encoder to embed the text blocks into a set of local representation vectors of all text blocks within the window. This set is used to help the model learn the relative position, alignment, and whitespace distribution pattern of the text blocks at the local level. The second stage is the global aggregator, which concatenates the local representation vectors of all windows according to the coordinate position of the windows on the page. It performs cross-window semantic integration through a standard multi-head self-attention mechanism (a total of 8 attention heads, 2 of which are specifically used to capture the lateral offset of text blocks to identify misaligned cases, and the offset > 3 pixels is defined as misaligned). Finally, it outputs the row boundary probability map and column boundary probability map with the same resolution as the input page, as well as a candidate list of cells containing the text block's affiliation. During the training phase, a multi-task combined loss is adopted: first, the binary cross-entropy of row and column boundary probabilities (positive samples are manually labeled row and column boundary pixels, and negative samples are non-boundary pixels); second, the semantic consistency loss within cells (based on the text block vectors output by the global aggregator, candidate cells are initially divided by the boundary probability map, and then the average cosine distance of all text block vectors in the same cell is calculated. The reciprocal of this average is used as the loss term to minimize the distance and achieve semantic consistency constraints). At the same time, a mixed layout enhancement strategy is introduced (character spacing ±2 pixel perturbation, line spacing ±5 pixel perturbation, random 10%-20% local cropping of the region, and random switching of 3 font styles) to improve the robustness of the model to layout changes in real shooting scenarios. In the inference phase, the input image is first preprocessed with text block detection consistent with the training process, then LORE outputs the results, and finally it is linked with a preset general table template (containing a common table structure template library with 3-10 columns and 5-20 rows) for matching: the matching degree of the number of cells between the candidate cell list and the template (weight 0.4) and the matching degree of text block alignment (weight 0.6) are calculated, and the template with a comprehensive matching score > 0.8 is selected. If the scores are all lower than 0.8, the boundary probability map is corrected based on the blank rate distribution of the candidate cells (areas with a blank rate > 50% are judged as row and column boundaries), and the cell set is finally determined.

[0026] Cell cropping and multi-engine OCR fusion (data extraction stage after wired / wireless table parsing): First, precise cropping is performed on each generated candidate cell. The effective text area within the cell is determined by the text block detection results. Non-text areas (such as residual borders, stains, and blank gaps) are avoided during cropping. If the cell is tilted (the angle between the text line direction and the horizontal direction is >3°), rotation correction is first performed based on the text line vector, and then the smallest bounding rectangle is cropped. The original resolution after cropping is preserved (no scaling is performed to avoid text stretching / compression), and the image is uniformly converted to grayscale (denoising parameters: 5×5 Gaussian filter) to ensure the consistency of input for each OCR engine. In the OCR stage, for multilingual, special symbol, and handwritten annotation scenarios in industrial documents, three types of engines are called in parallel: open source PaddleOCR (loading industrial scenario pre-trained model + handwritten text-specific model), Tesseract (enabling multilingual package + handwritten recognition mode), and commercial OCR service (configuring industrial symbol recognition plugin). Each engine outputs a structured result containing "character-by-character confidence, candidate text, and text length". The fusion phase employs a three-level weighted strategy: First, character position alignment is performed on the outputs of each engine (if the text length difference is greater than 1, the longest text is used as the baseline, missing character positions are filled with "empty characters" and the confidence level is set to 0). Then, three normalized indicators are calculated (all mapped to [0,1]): First, the character-level confidence average (the average confidence level of multiple engines is calculated position by position, and then the overall average of the cells is taken, with a weight of 40%). Second, the language model score (short text / special symbols use the n-gram model to calculate the inverse of perplexity, and multilingual long texts use the fine-tuned industrial scenario BERT model to output the text reasonableness probability, with a weight of 30%). Third, the template context consistency score (based on the field attributes of the preset table template, such as date fields matching the "year-month-day" format get 1 point, numeric fields without letters get 1 point, and non-matching results in a linear deduction of points according to the degree of deviation, with a weight of 30%). The three indicators are weighted and summed to obtain the fusion confidence level. For handwritten annotation scenarios, the automatic adoption threshold is lowered (from the usual 0.8 to 0.65, because the confidence level of handwritten text OCR is generally low). If the fusion confidence level is greater than or equal to the corresponding threshold, the result is automatically adopted; if it is less than the threshold, the field type is first determined: numeric / date fields trigger IoT two-way verification (comparing real-time data collected by IoT devices with the OCR result, such as temperature fields allowing ±0.5℃ error, date fields need to match exactly, and if the error exceeds the range, it is marked for manual review), and non-numeric / date fields are directly marked for manual review.

[0027] Field semantic validation and template matching are performed after OCR. Each template in the template library contains field location information, field semantic tags (such as "product number" and "temperature detection value"), field priority (high / medium / low, with high priority fields being core information such as product number and batch number), validation rules (regular expressions: such as product number must match "PRO-\d{6}", numerical range: such as temperature detection value must be between -20℃ and 80℃, unit constraints: such as pressure field must include "MPa" and "kPa" units), and layout vector (the layout encoder encodes the visual structure of the template, such as the field arrangement order and spacing ratio, into a 64-dimensional vector). Template matching consists of two steps: "layout similarity matching" and "field-level consistency verification." In the layout matching stage, the cosine similarity (60% weight) between the OCR-recognized field layout vector and the template layout vector is first calculated. Then, using the center coordinates of the template field as a reference, the Euclidean distance between the center of the recognized field and the center of the corresponding template field is calculated and converted to positional overlap (40% weight) according to the following criteria: 1 point for distance ≤ 5 pixels, 0.5 points for 5-10 pixels, and 0 points for > 10 pixels. The two are then weighted and summed to obtain a comprehensive layout matching score. Templates with a score ≥ 0.8 are selected as candidate templates. During the segment-level consistency verification stage, high-priority fields (product number, batch number, and inspection conclusion) of candidate templates are verified first. Format verification (e.g., the batch number needs to match the template regular expression "BATCH-\w{4}-\d{8}") and unit consistency verification (e.g., if OCR recognizes "pressure 2.5", the template's preset unit "MPa" needs to be added; if the recognized unit is "Pa", it is considered inconsistent). If necessary, the joint animal evidence collaboration module performs reverse verification (e.g., the recognized batch number is compared with the UID read from the RFID tag and the batch production record in the MES system, and they must be completely consistent). A template is marked as "passed" only when it simultaneously passes both layout matching (score ≥ 0.8) and field semantic consistency verification (100% compliance for high-priority fields, and ≤ 1 non-compliance item for medium- and low-priority fields). If the overall layout is similar (score 0.6~0.8) but some fields are inconsistent (1 non-compliance item for high-priority fields, or 2~3 non-compliance items for medium- and low-priority fields), it is marked as a partial match, triggering simultaneous manual review (prioritizing the correction of inconsistent fields) and active learning (adding the partial match cases to the template update dataset and iteratively optimizing the layout encoder and field matching rules). If the layout score < 0.6, it is directly marked as a non-match and template matching is re-initiated (expanding the template retrieval scope). The template matching results (passed / partially matched / non-matched), along with the OCR confidence scores of each field and verification logs, serve as the final data output credentials, packaged into an evidence package containing timestamps and verification node information for subsequent on-chain evidence storage (ensuring data immutability).

[0028] Output Format and Evidence Package: The final output includes three formats: standardized JSON, Excel spreadsheet, and HTML visualization. Specific rules are as follows: JSON Output: Includes field names, field values, joint confidence score (calculated by weighting OCR fusion confidence (60% weight), template matching score (30% weight), and field semantic verification score (10% weight), normalized to [0,1]), cell coordinates (represented by the coordinates of the top-left / bottom-right four points of the original image [x1,y1,x2,y2], rounded to one decimal place, generated through perspective matrix / rotation angle backprojection during preprocessing), template ID, template matching score (normalized to [0,1]), the OCR engine used and the average confidence score at the field level for each engine, and on-chain hash digest (the total hash of the evidence package, generated based on the SHA-256 algorithm); Excel Export: Maps the fields and columns of the template to Excel cells (preserving the row and column spacing and order of the template layout), including three sheets: original recognition results, review results, and evidence hash. All sheets include an audit column (review comments: enumerated values ​​"Pass / Correct / Reject", reviewer: employee ID format, review time: YYYY-MM-DD HH:MM:SS). HTML visualization output: Adapted for front-end manual verification, with wired tables overlaid with U-Net line probability maps, and wireless tables overlaid with LORE attention layers (supporting layer show / hide switching). Recognized fields are highlighted according to confidence level (≥0.8 green, 0.6-0.8 yellow, <0.6 red). Clicking on a highlight will jump to the corresponding coordinate position in the original image. The evidence package includes the original image, preprocessing transformation logs, intermediate feature maps of U-Net / CycleCenterNet / LORE (such as line probability maps and attention maps), OCR engine output and fusion records, template matching records and verification logs. The evidence package also includes key data hashes: original image hash. Structured data hashing IoT verification result hash Template matching hash These hashes are used in subsequent writing to the trusted evidence storage module to ensure that the identification process and results are immutable and traceable.

[0029] In this embodiment, the physical evidence collaboration module: First, identification resolution is performed: At workstations or warehouse nodes, the system periodically scans tag carriers using multi-mode IoT readers. These carriers may be RFID tags, NFC embedded chips, or QR codes / DM codes. For RFID and NFC, the system reads EPC, TID, user area extension area, and timestamp from the physical layer. It then performs an identity reliability score based on CRC results, RSSI signal strength, and antenna reception phase difference to eliminate misreading caused by environmental noise and reflections. For QR code carriers, image quality estimation and BCH verification are performed after reading to ensure the code data has not been manually altered. All identification data is uniformly encoded into a standardized physical evidence UID, while retaining physical layer characteristics (such as RSSI curves, PUF fragments, and QR code timing decoding time) as unforgeable evidence for subsequent trusted storage.

[0030] After the identifier is obtained, the module performs multi-source matching on the physical evidence UID, matching it with the field sets output by the production, warehousing, and document parsing modules respectively. For production matching, the system retrieves the production record corresponding to the UID from the MES / PLC / industrial control data bus, including production batch number, material number, equipment number, material loading time, workstation process parameters, first inspection / inspection records, etc., and sorts them chronologically. To ensure that this process is valid across different enterprises, the system calculates a hash for each process record and concatenates them sequentially to form a production chain summary. ,in For a safe hash function, This indicates concatenation. This summary is combined with the structured field summary output by the document parsing module. The data will be merged into a cross-stage trusted chain during subsequent on-chain evidence storage. Production records are used at this stage to reverse verify certificate fields. For example, when the document parsing module identifies "product number, batch number, inspection items, and manufacturing date," the system will perform consistency calculations on these fields with the MES approval log and real-time parameter records uploaded by the equipment. Strict equality checks are used for critical fields, and interval checks are used for time fields (e.g., the manufacturing date must be later than the completion time of all processes and not exceed the system's allowed buffer period). Inconsistent results will trigger a decrease in confidence and be written into the evidence package as an abnormal event.

[0031] Warehouse matching follows production matching. The module connects to the WMS and warehouse sensor network to collect real-time data on temperature, humidity, vibration, tilt, lighting, and access control status, binding this data to the actual location of the product's UID. The data stream from the warehouse sensors is segmented into small windows based on UID granularity. Each window calculates an environmental compliance score, the core function of which is: ,in This refers to the sensor readings within the window. Spec represents the product's permissible range; exceeding this range incurs penalties. The environmental scoring and document parsing modules directly bind to the "storage conditions, expiration date, and inspection conclusion" fields to verify the consistency between the storage conditions claimed on the certificate of conformity and the actual warehousing environment. For certain products that are extremely sensitive to the environment (such as chemicals and medical supplies), the module also compares small fluctuation patterns over a short period, such as the variance of adjacent 10-minute windows, as an indicator of whether environmental disturbances are abnormal, thereby further improving fraud detection capabilities.

[0032] After data binding is completed, the system begins evidence consistency verification. The module compares the four fields—evidence UID, production chain record, and warehousing environment chain record—according to their semantics, constructing a consistency score for each field. Taking batch number as an example, the consistency score is as follows: if the batch number output by OCR is exactly the same as the MES record, the score is 1; if one of the OCR multi-engine candidate values ​​matches the MES, the score is reduced proportionally according to the confidence level; if the semantics of the field determined by the text template (e.g., the batch number should be numbers + letters + year, etc.) are consistent with the MES format, but the specific value is inconsistent, the score is 0 and it is marked as a high-priority anomaly. Verification of the manufacturing date and process time uses a time consistency function: if... The score is then calculated using the distance decay function. ,in This indicates that offsets are allowed. All field scores are combined to form a total consistency score, which is then sent back to the document parsing module to adjust field confidence and output status, thus creating a closed loop between the two modules.

[0033] The evidence collaboration module performs abnormal behavior detection, using time series analysis to monitor RFID signal strength patterns, sudden changes in the storage environment, and abnormal reading frequencies. For example, if an RFID tag experiences an abnormal transition in a short period of time (signal noise pattern is different from the actual movement pattern), the system will determine whether there is tag removal or evidence replacement. If the warehouse suddenly darkens briefly in a bright environment, it may indicate that someone is blocking the camera or trying to manipulate the tag. These anomalies are written into the evidence package as events and trigger linked business rules, such as preventing entry, prohibiting release, and initiating manual review.

[0034] Ultimately, all data generated by the physical evidence collaboration module (including UID parsing, production chain summary, warehouse chain summary, field consistency score, and anomaly events) is merged and packaged with the structured data provided by the document parsing module to generate a physical evidence-document joint evidence package. This evidence package contains six core hash digests: original image hash, structured data hash, IoT UID hash, production chain summary hash, warehouse chain summary hash, and field consistency summary hash, which are used for writing to the consortium blockchain for evidence storage. Since each hash comes from a real physical event or model output and carries a timestamp, reader identifier, and sensor identifier, the entire certificate-physical link constitutes a legally verifiable, traceable, and non-repudiable digital credential.

[0035] During operation, the module feeds back to the document parsing module. If the fields identified by the model in low-quality documents or new templates are judged to be accurate through IoT verification, they are marked as high-confidence samples and sent to incremental training. If a situation occurs where "the model has high confidence but IoT verification fails", it is judged as a misidentified sample and enters the high-weight misidentified case pool, prompting the model to prioritize the optimization of this type of structure in the next training.

[0036] In this embodiment, the trusted evidence storage module: After document parsing, the module produces structured data, template matching results, field confidence chains, model version records, and processing logs. The evidence collaboration module provides evidence UIDs, production chain summaries, warehousing chain summaries, physical layer acquisition evidence (such as RFID TID bits, RSSI curves, and PUF local features), consistency scores, and anomaly events. The trusted evidence storage module receives this high-dimensional data from both modules and first performs evidence normalization, normalizing all fields according to the evidence semantics of identity-process-environment-behavior categories, compressing multi-source data into hash groups that can be expressed on-chain. To ensure the immutability of this transformation process, the module constructs a strict hash chain, where each piece of evidence's hash includes structured sorting, field value hashes, field confidence summaries, model version strings, and timestamps, which are concatenated to generate the total evidence hash. For example, a document evidence summary can be represented as... ,in The hash of the field value, For document parsing model version hash, To identify timestamps, the evidence chain digest is constructed similarly, but based on UID, production chain hash, storage chain hash, and physical layer feature hash, thereby ensuring that document evidence and physical evidence are fundamentally unforgeable.

[0037] After evidence preparation is complete, the system enters the cross-source evidence association phase. The module uses the overall consistency score generated in the physical evidence collaboration phase as the core reference, but before storage, it performs secondary calibration on key fields. For example, it compares the batch number hash parsed from the document with the batch number hash in the production chain summary; it calculates the semantic distance between the product model field in the document and the MES model field; and it performs consistency sampling calculations (such as window-level statistical distribution differences) between the valid date field in the document and the temperature and humidity fluctuation pattern in the warehousing chain. If these cross-source calibrations show anomalies, the trusted evidence storage module will not directly refuse to write, but will record an anomaly marker and write it into the evidence package. Subsequently, it will allow the writing of records with anomaly markers on the chain, providing an evidence chain for subsequent audits.

[0038] After evidence association is completed, the module converts all evidence into a multi-layered hash ledger structure writable by the consortium blockchain. This consortium blockchain is not a traditional single chain, but rather employs a three-layer logical ledger to adapt to industrial collaborative environments: the identity layer ledger records the binding relationship between product entities (UIDs), certificate numbers, and enterprise entities; the process layer ledger records the temporal behavior in document parsing and physical evidence collaboration, including identification events, reading events, sensing events, template matching events, and abnormal events; the compliance layer ledger records all field validations, consistency calculations, model version switching, and proactive learning trigger records, enabling hierarchical identification of responsibility points during regulatory or cross-enterprise tracing. To ensure logical traceability and non-duplication in storage, the system adopts a hierarchical hash referencing mechanism: each record in the process layer references the UID hash in the identity layer, and each record in the compliance layer references the event hash in the process layer, ensuring a single and unbranchable audit path through chain nesting.

[0039] The consortium blockchain employs a credit-weighted consensus mechanism, combining enterprise credit stratification, device data collection credibility models, and node behavior history to determine each node's write voting power. Node weights. It consists of three parts: enterprise reputation weight, equipment credibility weight, and historical behavior weight. The comprehensive score formula can be expressed as follows: ,in Based on the enterprise's historical reputation in the consortium blockchain (such as the number of violations and the audit pass rate). Based on the quality of equipment data acquisition (false reading rate, noise rate). Based on the behavior records of nodes participating in the consensus process.

[0040] The trusted evidence storage module constructs a visual smart contract layer. The smart contracts are responsible for verifying the format and field compliance of evidence and executing a cross-node visual execution graph on the blockchain, ensuring that every write can be traced back to an event graph. This includes information such as the triggering node, evidence root hash, associated event hash, model version, timestamp, and highlighted anomalies. The smart contracts also verify causal relationships between evidence: for example, a UID's manufacturing event must precede a warehouse entry event; a field inconsistency event must precede downstream actions; and if an anomaly event is not manually verified, the product will automatically be labeled with a risk in subsequent stages of the chain.

[0041] After being written to the consortium blockchain, the trusted evidence storage module provides bidirectional evidence collection capabilities both on and off-chain. On-chain storage includes hashes and minimum necessary fields, while the off-chain evidence repository preserves complete intermediate data and processing logs, including attention maps, line probability maps, cropping coordinates, model versions, OCR engine confidence records from the document parsing module, and production chain, storage chain, and physical layer data from the physical evidence collaboration module. Complete off-chain evidence can be quickly located using on-chain hashes, enabling high-completeness and high-reliability evidence collection in regulatory or judicial scenarios, ensuring the legal validity of the identification process.

[0042] The module employs high-performance, lightweight blockchain node technology in its engineering deployment, achieving millisecond-level latency for single-record writes, thus meeting the real-time requirements of production lines and warehouse inbound / outbound operations. Consortium blockchain nodes can be deployed across different enterprise sides, regulatory bodies, and the cloud, with encrypted P2P networks used for inter-node communication to ensure secure data transmission. To ensure long-term maintainability, the module permanently archives every model upgrade, template update, threshold adjustment, and business rule revision as an on-chain configuration snapshot, ensuring that system behavior at any given time can be rigorously reproduced and guaranteeing a closed audit chain.

[0043] In this embodiment, the business linkage module: First, an interpretable business semantic context is constructed: During the semantic context assembly process, the system extracts core tags for business identification from the structure fields output by the document parsing module, such as product number, batch number, production date, inspection conclusion, and expiration date, and maps them to the ERP material master data model, the quality management system's inspection master file, and the MES process chain structure. Physical UIDs, production chain summaries, and warehouse environment consistency records from the physical evidence collaboration module are mapped to "physical state vectors," representing the product's lifecycle position in the physical world and serving as input for business decisions. For example, if a physical evidence UID shows multiple abnormal fluctuations in warehouse temperature and humidity sensor records, the vector will automatically carry a risk weight; if the production chain summary indicates that the batch of products has parameter offsets in a certain process, the "production consistency tag" will be marked as unstable. The on-chain address and compliance tag generated by the trusted evidence storage module become "business legitimacy anchors," guiding the process system to distinguish between "trusted events," "abnormal events," and "events pending review." When integrating this data, the business linkage module reconstructs the business logic network according to the causal reference relationship in the trusted chain, so that the identification event, IoT behavior, verification behavior and evidence storage behavior enter the process system as a complete link.

[0044] After semantic construction is completed, the business linkage module will convert it into automatic triggering conditions, so that the business flow can proceed automatically based on the evidence status without human intervention. For example, when the document parsing module indicates that the template matching is completely successful, the field confidence is high and stable, the consistency score obtained from the OCR result and the physical evidence collaboration is close to full, and the hash groups of the three modules have been successfully uploaded to the blockchain in the trusted evidence storage module, the system will automatically generate an "allow entry" event and synchronize this event to the WMS through a plug-in interface, so that the warehouse can perform the entry action without manually checking the paper certificate of conformity. If the confidence of some fields is low, but the IoT verification shows that the physical object and the document are basically consistent, the system will generate a "sampling inspection required for entry" status in the WMS, so as not to delay circulation and ensure quality stability. If the results of the two modules contradict each other, for example, the document parsing field shows that the batch number is A, the physical evidence UID corresponds to the batch number B in the MES process chain, and the Trusted Ledger generates a "cross-source inconsistency" mark in the verification layer ledger, the business linkage module will push the entire business item into the "freeze process", block entry, prohibit exit, and issue a "high-risk batch" event to the ERP.

[0045] During the interconnection of business systems, the module deeply integrates with the enterprise's existing ERP / MES / WMS systems through standardized business event plugins. The plugins use an event-driven model, treating changes in the state of credible evidence as business triggers. For example, upon receiving an event stating "Certificate verification completed and physical evidence matched," the ERP system automatically records the material. Upon receiving an event stating "Physical evidence consistency anomaly in this batch of products," the MES system automatically locks all incomplete work orders for that batch and adds a quality inspection step. Upon receiving an event stating "Warehouse environment does not meet the certificate of conformity declaration conditions," the WMS system automatically refuses to change the location of the material. Each event carries a reference address to a credible evidence chain, enabling the business system to reassess risks when necessary by querying on-chain and off-chain evidence, thus making process execution not only automated but also interpretable.

[0046] To maintain system stability and long-term operational capability, the business linkage module continuously monitors the quality and behavior patterns of the evidence chain during execution. It automatically adjusts business decision thresholds using confidence distributions from the document parsing module and consistency trends from the physical evidence collaboration module. If OCR confidence continuously declines over a period, but RFID collection and production chain consistency remains stable, the system automatically assumes "the identification end is affected by the environment, but the physical end is reliable," thus appropriately lowering the document confidence threshold to ensure uninterrupted business operations. If IoT consistency fluctuates, the system automatically tightens the threshold to enhance control. This mechanism is not independent of business rules but relies on the statistical attributes of the reliable evidence chain layer by layer, making the decision-making mechanism adaptive.

[0047] Furthermore, when executing each business action, the business linkage module writes the execution record, triggering reason, cited evidence hash, on-chain event ID, and human involvement status back to the compliance layer ledger via the trusted evidence storage module. This keeps the process minutes synchronized with the evidence chain, preventing actual business execution from deviating from the data automation system and creating an off-chain black box. In this way, whether it's a quality dispute, a goods dispute, a process dispute, or a regulatory inspection, a consistent causal path can be formed through the on-chain ledger and the off-chain evidence repository, thereby achieving an irrefutable business supervision structure.

[0048] In the system logic, this module integrates the causal relationships between events into the process engine, so that the occurrence of a certain event will automatically trigger downstream business actions. For example, after the upstream supplier completes the verification of the certificate of conformity and completes the physical evidence matching and writes it into the consortium blockchain, the downstream enterprise's BPM Sync can automatically listen to the changes in the ledger and immediately generate a "pending receipt" event without manual notification; if the warehouse sensors record temperature and humidity fluctuations exceeding the standard for two consecutive hours, even if the document parsing and physical evidence collaboration are normal, the system will automatically trigger the action of "re-verification of the validity of the certificate of conformity is required", so that the process has the ability to proactively perceive risks; if there is a process deviation in the production chain record of a certain batch, the system will automatically insert an additional quality inspection node in the future outbound process of that batch, so that the business path dynamically changes with the changes in the credible evidence chain.

[0049] In supply chain collaboration, once a company completes the verification of certificates of conformity and matching of physical evidence, the event is simultaneously entered into a trusted ledger and automatically notified to partners. Partners can decide whether to accept goods based solely on the on-chain markers, eliminating the need for repeated identification and significantly reducing the cost of redundant inspections while improving supply chain responsiveness. Because all business actions are verifiable through a hierarchical hash structure within the trusted evidence storage module, downstream companies do not need to trust the upstream system itself but rather the on-chain evidence, thus achieving a cross-organizational trust-minimized business flow.

[0050] In this embodiment, the visualization governance module: The document parsing module provides field confidence chains, segmentation probability graphs, attention graphs, multi-engine OCR results, and template matching logs; the physical evidence collaboration module provides RFID signal curves, UID trajectories, production chain summaries, warehouse environment change sequences, consistency scores, and abnormal behavior records; the trusted evidence storage module provides event hashes, structured reference relationships, and smart contract execution records for the three-layer ledger; the business linkage module provides all automatically triggered business events, system threshold change records, manual review trajectories, and process decision chains; and the visualization governance module deconstructs these pieces of evidence from different dimensions into composable governance primitives.

[0051] After obtaining these governance primitives, the system constructs a multi-track visual timeline, connecting the chain of "document recognition → field generation → physical evidence scanning → consistency verification → evidence storage verification → business actions" into an event trajectory according to event time, source node, and reference relationship. This allows managers to view the entire lifecycle evidence chain of a certificate of conformity from the moment it is photographed until the business is completed in an expanded / collapsed manner. For example, from the moment a certificate of conformity is photographed to its recognition, every confidence fallback, boundary heatmap prompt, and template judgment switch in the document parsing module will appear in the trajectory; subsequently, the line graph of physical evidence features formed by the signal strength and antenna phase difference when the RFID UID is read will naturally be appended to the trajectory; when the production chain summary is registered in the Trusted Ledger and a hash reference is generated, each step of the verification logic in the smart contract execution graph will appear as a node in the trajectory; finally, the process trajectories of automatic warehousing, freezing, and review by the business linkage module constitute the tail of the event chain.

[0052] Beyond the event trajectory, the module constructs an entity relationship view, presenting the relationships between product entities, certificate entities, UID physical entities, production equipment, warehousing nodes, business events, and on-chain events in a graph-like format. Each node in the graph carries core evidence from previous modules. For example, document nodes contain field confidence, layout vectors, and model versions; physical evidence nodes contain RSSI curves and warehousing environment feature vectors; on-chain nodes contain hash digests and contract execution logs; and business nodes contain operator information, trigger time, and risk level. Graph edges are formed by causal references, hash references, and process triggering relationships, enabling managers to intuitively identify anomaly clusters, risk propagation paths, and process bottlenecks from the graphical structure. For instance, if a batch of products experiences multiple process fluctuations at a certain workstation, that workstation node will show an anomaly area with "increased event density" in the graph; if a type of certificate identification experiences a systematic decrease in confidence under certain lighting conditions, the "low-confidence chain" between the document node and the shooting environment node will become apparent; if the temperature and humidity in a warehouse area frequently become out of control, the warehousing node will form a high-frequency anomaly edge in the graph. This graph-based governance, based on the chain of evidence, frees quality risks from data tables and logs, making them visible on the management interface.

[0053] The visualization governance module further incorporates all visible nodes into the governance intervention mechanism, allowing managers to directly modify processes, thresholds, and strategies through the interface. Because all process thresholds, risk weights, and triggering logic of the business linkage module can be configured off-chain and ultimately written to the trusted evidence storage module to generate version snapshots, the governance module provides a traceable threshold adjustment view on the interface. For example, when a manager observes a consistency fluctuation pattern in the production chain of a certain workstation in the visualization interface, they can directly increase the quality weight of that workstation or decrease the pass-through threshold in the visualization governance interface. The modified strategy is then executed synchronously through the business linkage module and solidified into the new version configuration in the ledger by the trusted evidence storage module, forming a closed loop of "governance interface—business strategy—on-chain evidence storage." Since all change operations are synchronously written to the chain, every configuration adjustment in the governance module is auditable and non-repudiable.

[0054] During long-term operation, the visual governance module maintains a "governance evolution chain," automatically recording all governance actions, strategy changes, system adaptive adjustments, anomaly density changes, and model version changes in chronological order as governance snapshots. This snapshot is linked to the Trusted Ledger's compliance layer ledger, enabling the system's governance capabilities to be traced, verified, and analyzed. Managers can observe changes in identification accuracy, evidence consistency, warehouse environment stability, and process risk conditions over the past year through the governance evolution chain, allowing for timely adjustments to management strategies or upgrades to equipment environments, thus achieving data-driven, sustainable, and evolvable governance processes.

[0055] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A certificate of conformity intelligent analysis and management system based on evidence collaboration, characterized in that, The system includes: a document parsing module, a physical evidence collaboration module, a trusted evidence storage module, a business linkage module, and a visual governance module, wherein: The document parsing module uses a multi-model table classification algorithm to determine the table type of the input certificate image; based on the determination result, it selects either a wired table structure parsing algorithm or a wireless table structure parsing algorithm to generate the table structure; and combines a multi-OCR fusion recognition engine with template verification to output the structured field content and field confidence of the certificate image. The physical evidence collaboration module reads RFID / NFC / QR code identifiers and collects multi-source sensor data, and performs cross-source consistency verification between the structured fields and the product's physical production chain, environmental chain, and signal feature sequences to form a consistency score. The trusted evidence storage module adopts a hierarchical hash evidence structure and a credit-weighted consensus mechanism to write field content, evidence verification results and business behavior into a trusted ledger using visualized smart contract rules. The business linkage module automatically triggers or blocks enterprise business processes based on consistency scores and on-chain compliance records. The visualization governance module presents the document parsing path, physical evidence verification link, on-chain evidence structure and business actions in the form of an event timeline and entity relationship diagram, forming a governance interface based on the evidence chain.

2. The system according to claim 1, wherein, The table classification fusion algorithm includes YOLO, PaddleClas, and Qanything, and outputs the final table type determination result with a confidence weighting strategy to improve the classification accuracy under complex lighting and deformation conditions.

3. The system according to claim 1, wherein, When processing wired tables, the document parsing module obtains candidate table line segments through the pixel-level confidence map output by U-Net, and constructs a cost function based on line segment direction consistency, endpoint spacing and confidence response. It then uses a graph theory connectivity reconstruction algorithm to generate a set of cells that satisfy rectangular topological constraints. The graph theory reconstruction uses minimum cost cycle or minimum cost flow search to complete broken line repair and table structure completion.

4. The system according to claim 1, wherein, The wireless table parsing adopts the LORE two-stage Transformer model. Its local stage is based on windowed self-attention to extract local visual features, and its global stage is based on self-attention mechanism to fuse row and column relationships and contextual semantics to infer cell boundaries, logical connections and table structure.

5. The system according to any of the preceding claims, wherein, The document parsing module employs a multi-OCR engine fusion, including at least two character recognition engines that output results in parallel, and generates a joint confidence score based on character confidence, language model score, and template semantic consistency; the joint confidence score serves as document-side evidence and participates in weight allocation in the subsequent consistency scoring.

6. The system according to any of the preceding claims, wherein, The evidence collaboration module includes: Physical unique identification reading unit based on RFID or NFC; A sensor data acquisition unit used to collect temperature, humidity, vibration, current, voltage, or process status. Feature vectorization unit used to extract RSSI, phase difference, and PUF randomness features of signal feature sequences; A scoring unit used to compare feature vectors with field semantics, production timelines with environmental sequences, and generate consistency scores based on a weighted matching function.

7. The system according to any of the preceding claims, wherein, The consistency score is a combination of field matching probability, time deviation decay function, environmental compliance comparison function and signal feature stability parameter, used to quantify the degree of correspondence between the certificate content and the actual product in the production chain and environmental chain dimensions.

8. The system according to any of the preceding claims, wherein, The trusted evidence storage module constructs a layered evidence structure consisting of an identity layer, a process layer, and a compliance layer. The evidence in each layer forms an evidence chain through hash digests and citation relationships. The module adopts credit-weighted consensus to improve verification efficiency in cross-enterprise scenarios and performs on-chain verification of evidence format integrity, field logical relationships, and business process constraints through visual smart contracts.

9. The system according to any of the preceding claims, wherein, The business linkage module maps the consistency scoring threshold, template matching results, and on-chain compliance status into business rules, and connects with enterprise ERP, MES, and WMS systems to automatically trigger business processes such as warehousing, release, freezing, and review; the module further writes business execution traces as process-level evidence into the trusted ledger.

10. The system according to any of the preceding claims, wherein, The visualization governance module constructs an event timeline and entity association diagram based on the hierarchical evidence structure, displaying the document parsing process, intermediate evidence, physical evidence behavior chain, on-chain verification status, and business actions; and supports visualized adjustment of the parsing model version, threshold parameters, or risk weights on the governance interface. The adjustment records are written into the trusted ledger as compliance layer evidence to support long-term, auditable governance evolution.

Images