Supply chain penetration monitoring method and system based on document engine

Abstract

This invention discloses a supply chain penetration supervision method and system based on a document engine, comprising: firstly, automatically acquiring heterogeneous document clusters across the entire supply chain and encoding them into document semantic vectors by deploying an integrated document engine; secondly, performing cross-document feature fusion on the semantic vectors to extract the core feature set of the supply chain; and thirdly, simultaneously performing two key calculations: predicting potential risks based on the core feature set across multiple regulatory dimensions to generate potential risk correlation indices for each dimension; and fourthly, analyzing the document semantic vectors to dynamically determine the regulatory priority coefficients for each dimension. Finally, fusing the potential risk indices and priority coefficients to obtain a comprehensive evaluation value, and automatically generating regulatory alarms, evaluation reports, or rectification lists to feed back to the terminal. This invention achieves automated, intelligent penetration identification and dynamic, precise supervision of supply chain risks.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, and more specifically, to a supply chain penetration monitoring method and system based on a document engine. Background Technology

[0002] As supply chains become increasingly globalized and complex, the need for effective, penetrating oversight is growing. Traditional supply chain oversight methods typically rely on manual verification of isolated reports or simple rule comparisons, making it difficult to achieve data correlation and integration across the entire supply chain. Due to the heterogeneity of information systems and inconsistent data formats across different links in the supply chain, severe data silos have formed, leading to hidden risk transmission paths and making it difficult to detect potential problems in advance. Existing technologies lack the ability to automatically extract deep semantic features from massive, multi-source, and heterogeneous document data and perform intelligent correlation analysis, failing to dynamically assess the urgency of different risk dimensions, resulting in delayed and inefficient regulatory decisions. Summary of the Invention

[0003] The purpose of this invention is to provide a supply chain penetration monitoring method and system based on a document engine.

[0004] In a first aspect, embodiments of the present invention provide a supply chain penetration monitoring method based on a document engine, comprising:

[0005] Deploy and launch a document engine for supply chain penetration supervision, which integrates a full-link document interaction module, a supply chain penetration supervision model, and a supervision decision output module;

[0006] The end-to-end document interaction module connects to the end-to-end data collection node of the supply chain supervision entity to obtain the end-to-end document cluster of the supply chain supervision entity;

[0007] Text vector encoding is performed on the entire document cluster of the supply chain supervision entity to obtain the document semantic vector corresponding to the entire document cluster;

[0008] Perform cross-document feature cross-matching and fusion processing on the semantic vector of the document to obtain the core feature set of the supply chain corresponding to the semantic vector of the document.

[0009] Determine at least one penetrating supervision dimension that corresponds to and is compatible with the full-link document cluster. Based on each penetrating supervision dimension, perform hazard prediction on the core feature set of the supply chain to obtain the hazard correlation index corresponding to each penetrating supervision dimension.

[0010] The document semantic vector is weighted and anchored according to the document semantic driving regulatory dimension to obtain the regulatory priority coefficient corresponding to each penetrating regulatory dimension;

[0011] Based on the hidden danger correlation index and the regulatory priority coefficient corresponding to each of the penetrating regulatory dimensions, the comprehensive evaluation value of the supply chain regulatory entity is determined.

[0012] The comprehensive assessment value is input into the regulatory decision output module of the document engine, which generates corresponding supply chain penetration supervision alarm information, compliance assessment report or rectification suggestion list, and feeds it back to the terminal device of the supply chain supervision entity through the end-to-end document interaction module.

[0013] In a second aspect, embodiments of the present invention provide a server system, the server system including a computer program, which, when running, controls the server system to execute the method described in the first aspect.

[0014] Compared to existing technologies, the beneficial effects of this invention include: A supply chain penetration supervision method and system based on a document engine, as disclosed in this invention, includes: First, by deploying an integrated document engine, automatically acquiring heterogeneous document clusters across the entire supply chain and encoding them into document semantic vectors. Next, cross-document feature fusion is performed on the semantic vectors to extract the core feature set of the supply chain. The method simultaneously performs two key calculations: predicting potential risks based on multiple regulatory dimensions of the core feature set and generating potential risk correlation indices for each dimension; and analyzing the document semantic vectors to dynamically determine the regulatory priority coefficients for each dimension. Finally, the potential risk indices and priority coefficients are fused to obtain a comprehensive evaluation value, and based on this, regulatory alarms, evaluation reports, or rectification lists are automatically generated and fed back to the terminal. This invention achieves automated, intelligent penetration identification and dynamic, precise supervision of supply chain risks. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the steps of the supply chain penetration monitoring method based on a document engine provided in this embodiment of the invention;

[0017] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0018] 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the supply chain penetration monitoring method based on a document engine provided in this embodiment of the present disclosure. The following is a detailed description of the supply chain penetration monitoring method based on a document engine.

[0021] Step S201: Deploy and start the document engine for supply chain penetration supervision. The document engine integrates a full-link document interaction module, a supply chain penetration supervision model, and a supervision decision output module.

[0022] Step S202: Connect the end-to-end data collection node of the supply chain supervision entity through the end-to-end document interaction module to obtain the end-to-end document cluster of the supply chain supervision entity;

[0023] Step S203: Perform text vector encoding processing on the full-link document cluster of the supply chain supervision entity to obtain the document semantic vector corresponding to the full-link document cluster;

[0024] Step S204: Perform cross-document feature cross-matching and fusion processing on the document semantic vector to obtain the supply chain core feature set corresponding to the document semantic vector;

[0025] Step S205: Determine at least one penetrating supervision dimension that is compatible with the full-link document cluster, and perform hidden danger prediction on the core feature set of the supply chain based on each penetrating supervision dimension to obtain the hidden danger correlation index corresponding to each penetrating supervision dimension.

[0026] Step S206: Assign weighted anchors to the document semantic vector based on the document semantic driving regulatory dimension to obtain the regulatory priority coefficients corresponding to each penetrating regulatory dimension;

[0027] Step S207: Based on the hidden danger correlation index corresponding to each penetrating supervision dimension and the supervision priority coefficient corresponding to each penetrating supervision dimension, determine the comprehensive evaluation value of the supply chain supervision entity;

[0028] Step S208: The comprehensive evaluation value is input into the regulatory decision output module of the document engine, and the document engine generates corresponding supply chain penetration supervision alarm information, compliance assessment report or rectification suggestion list, and feeds it back to the terminal device of the supply chain supervision entity through the full-link document interaction module.

[0029] In this embodiment of the invention, for example, after receiving the supply chain penetration supervision task instruction, the server first loads the pre-trained and configured "document engine" software module from the model repository and deploys it in the computing resource pool to start running. This document engine is an integrated intelligent processing system, its core integrating three functional modules: a full-link document interaction module, a supply chain penetration supervision model, and a supervision decision output module. The full-link document interaction module is responsible for secure and standardized data communication with various external enterprise information systems; the supply chain penetration supervision model is a composite artificial intelligence model containing multi-layer neural network components, used for deep analysis and decision-making; the supervision decision output module is responsible for converting the model's analysis results into business-readable instructions and reports. After the server completes the engine deployment, each module enters a standby state, waiting for data processing tasks.

[0030] Taking the supply chain monitoring of a new energy vehicle manufacturing company as an example, the server actively connects to various key data collection nodes in the company's supply chain through a full-link document interaction module, according to preset protocols and interfaces (such as API, database connection, file transfer, etc.). These nodes are distributed throughout the entire supply chain, including the upstream Supplier Management System (SRM) for collecting purchase orders, supplier qualification documents, price agreements, etc.; the company's internal Enterprise Resource Planning (ERP) and Manufacturing Execution System (MES) for collecting production work orders, material requirements planning, receiving slips, quality inspection slips, etc.; and the downstream Warehouse Management System (WMS) and Logistics Management System (TMS) for collecting outbound slips, logistics waybills, and in-transit tracking data. The server pulls all relevant electronic and image document data within a specified time window (such as the past quarter) from these heterogeneous data sources in a concurrent or asynchronous manner, forming a complex and diverse "full-link document cluster". This document cluster can contain API data in JSON format, database table records, scanned copies of contracts in PDF format, and signed receipt photos in JPG format, etc.

[0031] After obtaining the original end-to-end document cluster, the server cannot directly input it into the model because its data format and semantic expression are inconsistent. Therefore, the server first calls the "Basic Document Feature Extraction Component" in the supply chain penetration monitoring model to perform heterogeneous data alignment. Specifically, the semantic mapper within this component begins to work. For example, for a PDF "Raw Material Purchase Order" from a supplier, the semantic mapper identifies its document attribute category as "Purchase Document," and then determines that the semantic feature domains to be extracted include "Material Code," "Purchase Quantity," "Unit Price," "Supplier Name," and "Agreed Delivery Date." For the "Agreed Delivery Date" feature, its attribute value domain may be pre-divided into partitions such as "Urgent (within 3 days)," "Normal (4-15 days)," and "Relaxed (more than 15 days)." The semantic mapper uses OCR and natural language understanding technology to parse the specific delivery date from the PDF text as "October 25, 2023," calculates it with the current date to obtain a delivery cycle of "7 days," and then maps it to the "Normal" partition, outputting the structured data "Delivery Date Feature: Normal." By performing this operation on every key text unit of all documents in the document cluster, the server transforms the originally unstructured document cluster into a standardized, machine-readable "standardized document dataset," which exists in the form of structured key-value pairs or vectors.

[0032] Subsequently, the server invokes the embedding layer within the basic document feature extraction component to perform text vector encoding on this standardized document dataset. The embedding layer converts each structured feature value (such as "Material Code: A001", "Supplier Name: XX Steel", "Delivery Date Feature: Normal") into a high-dimensional, dense numerical vector through table lookup or neural network calculation. Finally, all feature vectors from all documents are concatenated or aggregated in a specific order (such as by timestamp and document type) to form a global "document semantic vector" that represents the semantic information of the entire supply chain document flow. This vector forms the basis for all subsequent in-depth analyses.

[0033] After obtaining the semantic vector of the document, the server inputs it into the "cross-document feature fusion sub-component" within the basic document feature extraction component. This sub-component is a deep neural network designed to discover and fuse scattered correlation features across different documents and stages. Continuing with the example of new energy vehicle manufacturing, this sub-component performs automatic cross-matching analysis: it correlates the features of "Supplier A" and "Purchased Materials - Battery Cells" in the "Purchase Order" with the features of "Supplier A" and "Batch Defect Rate" in the "Quality Inspection Report"; simultaneously, it correlates the feature of "Used Materials - Battery Cell Batch Number" in the "Production Work Order" with the features of "Transportation Destination" and "Temperature in Transit" in the "Logistics Waybill." Through the nonlinear calculations of a multi-layer neural network, this sub-component can fuse this cross-document correlation information to extract a higher-level, regulatory-significant "supply chain core feature set." For example, output features might include: "Quality stability index of a specific supplier - key materials," "Average turnover cycle of key materials from procurement to production," and "Timeliness and risk coefficient of a specific transportation route," etc. These characteristics are no longer simply a list of original documents, but rather deeply refined indicators that directly point to the health status of the supply chain.

[0034] Based on the current regulatory policies and the business scope involved in the document cluster, the server automatically determines the "penetrating regulatory dimensions" that need to be assessed. For example, for this automobile manufacturing company, the server might identify four core dimensions: "financial compliance risk," "product quality risk," "logistics fulfillment risk," and "supply chain resilience risk." Subsequently, the server calls the "specific assessment feature extraction component," which is sequentially connected in the supply chain penetration regulatory model. This component contains four independent sub-models, each corresponding to one of the four dimensions mentioned above. The server then inputs the "supply chain core feature set" obtained in the previous step into these four specific assessment models simultaneously.

[0035] Each specialized model performs in-depth risk prediction for its respective regulatory dimension. For example, the "Product Quality Risk" model focuses on quality-related features in the "Core Feature Set," such as "historical trends in battery cell batch defect rates" and "differences in quality inspection standards among different suppliers." It calculates a "Risk Correlation Index" between 0 and 1 using its internal algorithm (such as a classifier or regressor), for example, 0.85. The higher the index, the greater the likelihood of a risk in that dimension. Similarly, the "Logistics Fulfillment Risk" model might output 0.60, the "Financial Compliance Risk" model 0.45, and the "Supply Chain Resilience Risk" model 0.70. At this point, the server obtains a multi-dimensional risk quantification snapshot.

[0036] The hazard correlation index reflects the absolute risk level of each dimension, but in actual supervision, the urgency of different risks varies at different times and under different business contexts. Therefore, the server calls the "weight allocation feature extraction component" in the supply chain penetration supervision model in parallel. This component takes the most basic "document semantic vector" as input and analyzes the overall semantic context of the current supply chain business. For example, by analyzing the document semantic vector, the server finds that the proportion of "urgent orders" in "purchase orders" has increased significantly recently, "production work orders" have tight scheduling, and "logistics waybills" show weather warnings for some routes. Based on this context, the weight allocation component dynamically generates "regulatory priority coefficients" for each regulatory dimension through algorithms such as attention mechanisms. It may determine that "logistics fulfillment risk" and "supply chain resilience risk" have the highest priority under the current business conditions, assigning them higher coefficients (e.g., 0.4 and 0.35), while "financial compliance risk" is relatively less important at this time, assigning it a lower coefficient (e.g., 0.1). The sum of these coefficients is 1, reflecting the key areas where regulatory resources should be allocated.

[0037] The server performs a weighted fusion calculation of the hazard correlation index for each regulatory dimension and its corresponding regulatory priority coefficient. A specific calculation method is: Comprehensive Assessment Value = Σ(Hazard Correlation Index of the i-th Dimension × Regulatory Priority Coefficient of the i-th Dimension). Substituting the example data above: Comprehensive Assessment Value = 0.45 + 0.1 + 0.85 + 0.15 + 0.60 + 0.4 + 0.70 + 0.35 = 0.045 + 0.1275 + 0.24 + 0.245 = 0.6575. This comprehensive assessment value of 0.6575 considers not only the absolute magnitude of the risk in each dimension (hazard correlation index) but also its relative importance in the current business context (regulatory priority coefficient), thus forming a more comprehensive and practical overall risk assessment score.

[0038] The server inputs the calculated comprehensive assessment value (0.6575) into the regulatory decision output module. This module has multiple preset decision rules. For example, the rule stipulates that when the comprehensive assessment value is greater than 0.6, a "yellow" alarm message is generated. At the same time, based on the risk index of each dimension, a detailed "compliance assessment report" is automatically generated. The report will clearly indicate that the "product quality risk" dimension index is high (0.85), and it is recommended to focus on reviewing the cell quality control process of supplier A; the "logistics fulfillment risk" dimension index is 0.60 under the current high priority, and it is recommended to activate the backup transportation route plan. The server may also generate a specific "rectification suggestion list" listing the pending items. Finally, the server pushes this regulatory decision package, which includes alarm information, assessment report and rectification list, to the terminal device (such as computer, large screen or mobile application) of the supply chain management personnel of the new energy vehicle manufacturing company in a secure manner (such as through the enterprise gateway) through the end-to-end document interaction module, completing a complete penetrating regulatory closed loop. Meanwhile, the feedback data from this regulatory exercise (such as whether the company has confirmed the alarm and whether the rectification measures have been implemented) as well as the newly added document clusters will be recorded by the server and used to continuously iterate and optimize the processing rules of the document engine and the supply chain penetration supervision model, making it smarter the more it is used.

[0039] In this embodiment of the invention, the text vector encoding process performed on the full-link document cluster of the supply chain supervision entity to obtain the document semantic vector corresponding to the full-link document cluster can be implemented through the following example.

[0040] Heterogeneous data alignment is performed on the full-link document cluster to obtain a standardized document dataset of the full-link document cluster;

[0041] The standardized document dataset is subjected to document semantic vector text vector encoding processing to obtain the document semantic vector of the standardized document dataset.

[0042] In this embodiment of the invention, exemplarily, when the server performs the step of "text vector encoding processing of the entire document cluster," the specific operations are as follows: First, the server performs heterogeneous data alignment to generate a standardized document dataset. The server calls the basic document feature extraction component in the supply chain penetration monitoring model, and its built-in semantic mapper processes the original multi-format document cluster. For example, for a PDF purchase contract from a supplier and a JSON work order record from the production system, the semantic mapper will identify the former as a "purchase" document and extract key text units such as "material specifications," "purchase quantity," and "contract unit price"; and identify the latter as a "production" document and extract units such as "production batch number" and "planned completion time." For the "contract unit price" unit, the semantic mapper partitions according to predefined attribute value ranges (such as "low price range," "normal range," and "high price range"), mapping the parsed specific amount "105.00 yuan" to the feature value of "normal range." Through this operation, the unstructured text of all documents is uniformly transformed into a structured "standardized document dataset", which is a series of normalized (feature, value) pairs.

[0043] Subsequently, the server performs text vector encoding on the standardized document dataset to obtain document semantic vectors. The server then inputs this dataset into the embedding layer of the basic document feature extraction component. This embedding layer is a trained neural network layer that converts each discrete feature value (such as "Material Specification: Ternary Lithium Battery", "Price Range: Normal") into a high-dimensional dense numerical vector (i.e., the embedding vector). Next, the server aggregates all feature embedding vectors from all documents according to a preset temporal and logical order (e.g., through concatenation or averaging), ultimately generating a unified, fixed-dimensional "document semantic vector." This vector comprehensively encodes the complete semantic information of the entire document chain, providing a numerical foundation for subsequent deep feature analysis.

[0044] In this embodiment of the invention, the step of aligning heterogeneous data of the end-to-end document cluster to obtain a standardized document dataset of the end-to-end document cluster can be implemented through the following example.

[0045] When the full-chain document cluster of the supply chain regulatory entity is obtained, a supply chain penetration regulatory model for conducting supply chain penetration regulatory assessment of the supply chain regulatory entity is obtained; the supply chain penetration regulatory model includes a basic document feature extraction component.

[0046] The basic document feature extraction component determines the document attribute category that corresponds to and is suitable for the full-link document cluster, and extracts the document text unit of the full-link document cluster based on the document attribute category;

[0047] The document text unit is semantically decomposed using the basic document feature extraction component to obtain the semantic decomposition result corresponding to the document text unit.

[0048] The basic document feature extraction component performs structured expression processing on the semantic decomposition results of the document text units to obtain the standardized document dataset of the full-link document cluster.

[0049] In this embodiment of the invention, for example, after the server loads the acquired full-link document cluster of the supply chain regulatory entity into memory, it then retrieves and instantiates the supply chain penetration regulatory model for this assessment task from the model repository. One of the core components of this model is the basic document feature extraction component.

[0050] Next, the server invokes the document classifier within the basic document feature extraction component to intelligently identify and classify the input end-to-end document clusters. For example, for an image initially identified through OCR, the classifier determines it as a "Purchase Order" attribute category based on its layout and keywords (such as "PurchaseOrder" and "Waybill Number"); for a JSON record read from the database, the classifier determines it as a "Warehouse Inbound Order" attribute category based on its data structure and field names. After determining the attribute category of each document, the server locates and extracts key document text units from the raw data based on predefined extraction rules bound to each category. Taking a "Purchase Order" as an example, the server will extract text units such as "Supplier: Shanghai Precision Components Co., Ltd.", "Material Number: SPC-1002", "Quantity: 500", and "Unit Price (including tax): 120.00 yuan".

[0051] Then, the server uses the semantic mapper in the basic document feature extraction component to perform deep semantic decomposition on each extracted document text unit. The semantic mapper internally maintains a value range partitioning knowledge base for different business domains. Taking the text unit "Tax-inclusive unit price: 120.00 yuan" as an example, the semantic mapper identifies it as the "purchase unit price" feature. It queries the knowledge base for the partitioning threshold set for the historical price of this material (e.g., "normal range" is 100-130 yuan) and maps the value 120.00 yuan to the semantic label "normal range". This "normal range" label is the semantic decomposition result obtained after decomposing the text unit.

[0052] Finally, the server performs structured representation processing on all the semantic decomposition results of the above documents. It combines "Supplier: Shanghai Precision Components Co., Ltd." with the decomposition result "Supplier Rating: A", "Material Number: SPC-1002" with the decomposition result "Material Type: Key Production Material", and "Tax-Inclusive Unit Price: 120.00 RMB" with the decomposition result "Price Range: Normal". All units of all documents undergo this processing, ultimately generating a unified, machine-readable, standardized document dataset. This dataset is typically organized in list or table format, with each record containing structured fields such as "Document ID", "Attribute Category", "Feature Fields", "Original Value", and "Standardized Semantic Value" to facilitate subsequent vectorized encoding.

[0053] In this embodiment of the invention, the basic document feature extraction component includes a semantic mapper; the document text unit includes at least one document text unit, and the at least one document text unit includes a target document text unit;

[0054] The step of performing semantic decomposition on the document text unit using the basic document feature extraction component to obtain the semantic decomposition result corresponding to the document text unit can be implemented through the following example.

[0055] Based on the document attribute category to which the target document text unit belongs, the semantic feature domain corresponding to the target document text unit is determined; the semantic feature domain includes multiple feature values, and each feature value is associated with a corresponding attribute value domain partition.

[0056] The semantic mapper determines the attribute value domain partition to which the target document text unit belongs, and the feature value corresponding to the attribute value domain partition to which the target document text unit belongs is determined as the document semantic decomposition result corresponding to the target document text unit.

[0057] In this embodiment of the invention, for example, the server first determines the corresponding semantic feature domain based on the document attribute category to which the target document text unit belongs. For instance, when the server processes a "logistics waybill," it identifies one of the target text units as "planned delivery time: 2023-11-05 14:00." Based on the attribute category "logistics waybill," the server determines the semantic feature domain corresponding to this text unit as "delivery timeliness." This feature domain predefines multiple feature values, each associated with a clear attribute value domain partitioning rule. Specifically, the "delivery timeliness" feature domain can include four feature values: "on time," "minor delay," "serious delay," and "extremely serious delay." The rules associated with them are: a difference of less than 2 hours between the planned time and the actual receipt time is considered "on time"; a delay of 2 to 24 hours is considered "minor delay"; a delay of 24 to 72 hours is considered "serious delay"; and a delay exceeding 72 hours is considered "extremely serious delay."

[0058] Next, the server uses the semantic mapper in the basic document feature extraction component to determine the specific attribute value range partition to which the target text unit belongs, based on the current actual business data. The semantic mapper obtains the current system time (assumed to be 2023-11-05 18:30) as a simulation of the actual receipt time and calculates the difference between it and the planned time "2023-11-05 14:00", resulting in a delay of 4.5 hours. Subsequently, the semantic mapper matches this value with predefined rules, determining that the 4.5-hour delay falls within the "2 to 24 hours" range, i.e., belongs to the "minor delay" attribute value range partition.

[0059] Finally, the server determines the feature value "slight delay" corresponding to this partition as the final semantic decomposition result for the target document text unit "Planned delivery time: 2023-11-05 14:00". This result "Delivery timeliness: slight delay" will be output for subsequent structured expression processing.

[0060] In this embodiment of the invention, the core feature set of the supply chain is obtained by cross-document feature fusion sub-component in the basic document feature extraction component performing cross-document feature cross-matching and fusion processing on the semantic vector of the document; the basic document feature extraction component belongs to a supply chain penetration supervision model used to conduct supply chain penetration supervision assessment on the supply chain supervision entity;

[0061] In this embodiment of the invention, the following implementation methods are also provided.

[0062] Obtain the first training corpus; the unlabeled document training samples included in the first training corpus are document training samples without labeled regulatory target values;

[0063] Obtain at least one self-supervised training objective for the feature extraction component of the initial basic document;

[0064] The initial basic document feature extraction component performs forward feature propagation on the unlabeled document training samples to obtain intermediate feature representations associated with each self-supervised training objective. Based on the intermediate feature representations, at least one error value corresponding to the at least one self-supervised training objective is determined; one self-supervised training objective corresponds to one error value.

[0065] A first target error value is determined based on the at least one error value. The model parameters of the initial basic document feature extraction component are optimized based on the first target error value. The optimized initial basic document feature extraction component is then determined as the basic document feature extraction component.

[0066] In this embodiment of the invention, for example, the server acquires the first training corpus. The server extracts massive amounts of real supply chain document data accumulated over the past three years from a historical business database, such as 1 million purchase order records, 800,000 logistics waybill records, and 600,000 warehouse receipt records. These records collectively constitute the set of "unlabeled document training samples," i.e., the first training corpus. These samples only contain raw document information and have not been manually labeled with any regulatory target values ​​such as "high risk" or "compliance," thus fully utilizing the unlabeled raw data.

[0067] Next, the server defines and obtains at least one self-supervised training objective for the initial basic document feature extraction component to be trained. These objectives aim to enable the model to learn effective feature representations from the structure of the data itself. For example, the server sets two key self-supervised training objectives: the first is a "document heterogeneous enhancement contrast training objective," which aims to enable the model to learn to recognize the consistency between perturbed but semantically unchanged documents and their original versions in the feature space; the second is a "model distillation-guided training objective," which aims to enable a lightweight student model (i.e., the initial basic document feature extraction component) to mimic the feature extraction capabilities of a pre-trained, more complex teacher model.

[0068] Then, the server performs forward feature propagation on the unlabeled document training samples through the initial basic document feature extraction component to calculate the error value corresponding to each training objective. Specifically, the server takes a batch of training samples, such as a batch containing purchase orders and corresponding logistics waybills. For the "document heterogeneous enhancement contrast training objective," the server first injects perturbations into the text units in this batch of samples, for example, simulating OCR recognition errors by changing "120.00 yuan" to "120.00 yuan," or randomly obscuring some fields, generating corresponding "perturbed document training samples." Subsequently, the server inputs the original samples and perturbed samples into the initial basic document feature extraction component. This component sequentially performs heterogeneous data alignment, text vector encoding (through its embedding layer), and cross-document feature cross-matching and fusion processing (through its cross-document feature fusion sub-component), ultimately outputting two "intermediate feature representations": one is the "original intermediate feature representation" based on the original samples, and the other is the "perturbed intermediate feature representation" based on the perturbed samples. The server calculates the cosine similarity between the two feature representations in the vector space and compares it with the ideal value (which should be highly similar) to obtain the error value corresponding to the contrast training target.

[0069] Simultaneously, for the "model distillation-guided training objective," the server inputs the "original intermediate feature representation" of the same batch of original samples into a subsequent, simple hazard assessment head (i.e., processing unit) to obtain a preliminary "training hazard assessment result." On the other hand, the server calls a pre-trained, parameter-rich teacher model to process the same batch of original samples, generating a more accurate "simulated hazard correlation index." The server calculates the error value corresponding to the model distillation-guided training objective by comparing the difference (such as mean squared error) between the assessment results output by the student model and the simulated index provided by the teacher model.

[0070] Subsequently, based on at least one of the aforementioned error values, the server determines a first overall target error value through weighted summation and other methods. The server then uses the backpropagation algorithm to calculate the gradient based on this first target error value and optimizes and updates the model parameters of the initial basic document feature extraction component (including the embedding layer, mapping rule parameters in the semantic mapper, and network weights of the cross-document feature fusion sub-component). The server repeats the above sampling, forward propagation, error calculation, and backpropagation optimization process until the model performs stably on the validation set. Finally, the server determines the optimized and converged model as the "basic document feature extraction component" for supply chain penetration supervision.

[0071] In this embodiment of the invention, the at least one self-supervised training objective includes a document field mask restoration training objective;

[0072] The step of performing forward feature propagation on the unlabeled document training samples through the initial basic document feature extraction component to obtain intermediate feature representations associated with each self-supervised training objective, and determining at least one error value corresponding to the at least one self-supervised training objective based on the intermediate feature representations, can be implemented through the following example.

[0073] The initial basic document feature extraction component is used to perform heterogeneous data alignment on the unlabeled document training samples to obtain the first sample standardized document dataset of the unlabeled document training samples; the first sample standardized document dataset contains the semantic decomposition results of the sample documents at multiple data unit locations.

[0074] Select the first data cell position to be masked from the multiple data cell positions corresponding to the first sample standardized document dataset;

[0075] Based on the mask identifier, the semantic decomposition result of the sample document at the first data unit position in the first sample standardized document dataset is processed by data unit masking to obtain the training mask dataset.

[0076] The initial basic document feature extraction component performs document semantic vector text vector encoding processing on the training mask dataset to obtain the masked training document semantic vector of the training mask dataset. The masked training document semantic vector is then subjected to cross-document feature cross-matching and fusion processing to obtain the masked intermediate feature representation corresponding to the masked training document semantic vector. The masked intermediate feature representation belongs to the intermediate feature representation.

[0077] By using a processing unit adapted to the training target of the document field mask restoration, the intermediate feature representation of the mask is subjected to mask field restoration processing to obtain the predicted document semantic decomposition result corresponding to the mask identifier in the training mask dataset.

[0078] Based on the semantic decomposition results of the sample documents at the first data unit position in the first sample standardized document dataset and the semantic decomposition results of the predicted documents, the error value corresponding to the document field mask restoration training target is determined.

[0079] In this embodiment of the invention, for example, the server obtains an unlabeled document training sample, such as a data pair containing a "purchase order" and its associated "logistics waybill". The server performs heterogeneous data alignment processing on this sample using an initial basic document feature extraction component. Specifically, its semantic mapper parses "Supplier: Shanghai Precision Components Co., Ltd." in the purchase order as the "Supplier Name" feature, mapping the result to "Supplier Identity: Core Supplier"; it parses "Material Number: SPC-1002" as the "Material Code" feature, mapping the result to "Material Type: Key Production Part". Simultaneously, it parses "Carrier: Xunda Logistics" in the logistics waybill as the "Carrier" feature, mapping the result to "Carrier Level: Grade A". All these mapping results together constitute a structured "First Sample Standardized Document Dataset". This dataset logically consists of multiple data unit positions, each storing a specific sample document semantic decomposition result. For example, position 1 stores "Supplier Identity: Core Supplier", position 2 stores "Material Type: Key Production Part", and position 3 stores "Carrier Level: Grade A".

[0080] Next, the server selects one of these multiple data unit locations as the "first data unit location" for masking according to a preset random masking strategy (e.g., with a 15% probability). Let's assume that location 1 was randomly selected this time, corresponding to the "Supplier Identity: Core Supplier" field.

[0081] Next, the server performs data cell masking. It uses a special identifier, "[MASK]", to replace the original content at the first data cell location in the first sample of the standardized document dataset. Thus, the data at location 1 changes from "Supplier Identity: Core Supplier" to "Supplier Identity: [MASK]". The data at other locations remains unchanged. In this way, the server obtains a "training mask dataset" containing the missing information.

[0082] Subsequently, the server inputs this training mask dataset into the initial basic document feature extraction component for forward feature propagation. The embedding layer within this component converts all structured features, including "[MASK]", into numerical vectors, generating a "mask training document semantic vector". This vector is then fed into a cross-document feature fusion sub-component. This sub-component infers and fuses information about the mask locations based on contextual information (such as known "material type: critical production part" and "carrier grade: A"), ultimately outputting a comprehensive "mask intermediate feature representation".

[0083] Subsequently, the server invokes a processing unit adapted to the training target corresponding to the masked data reconstruction of the document fields. This unit can be a lightweight classification layer (or prediction head). This processing unit receives the intermediate feature representation of the mask and specifically predicts the masked location (the location of the first data unit). Based on the contextual patterns learned by the model, it calculates the probability distribution of all possible feature values ​​(such as "core supplier," "general supplier," "new supplier," etc.) and outputs "core supplier," which has the highest probability, as the "predicted document semantic decomposition result."

[0084] Finally, the server calculates the error value to guide model optimization. It compares the predicted result "core supplier" with the original true value (i.e., the "semantic decomposition result of the sample document") at the first data cell position in the first sample standardized document dataset. Since this is a classification task, the server typically uses the cross-entropy loss function to calculate the difference between the two; this difference value is the "error value corresponding to the document field mask restoration training target." This error signal propagates back through the model, driving it to adjust its parameters so that it can more accurately restore the masked key business fields according to the context in the next iteration. Through this type of training on massive amounts of samples, the model gradually masters the rich semantic logical relationships within the documents.

[0085] In this embodiment of the invention, the at least one self-supervised training objective includes a document anomaly field detection training objective;

[0086] The step of performing forward feature propagation on the unlabeled document training samples through the initial basic document feature extraction component to obtain intermediate feature representations associated with each self-supervised training objective, and determining at least one error value corresponding to the at least one self-supervised training objective based on the intermediate feature representations, can be implemented through the following example.

[0087] The initial basic document feature extraction component is used to perform heterogeneous data alignment on the unlabeled document training samples to obtain the first sample standardized document dataset of the unlabeled document training samples; the first sample standardized document dataset contains the semantic decomposition results of the sample documents at multiple data unit locations.

[0088] Select the second data unit location to be tampered with from the multiple data unit locations corresponding to the first sample standardized document dataset;

[0089] Based on the abnormal field values, the semantic decomposition results of the sample documents at the second data unit position in the first sample standardized document dataset are subjected to field tampering processing to obtain the training tampered dataset; the abnormal field values ​​are different from the semantic decomposition results of the sample documents at the second data unit position;

[0090] The initial basic document feature extraction component performs document semantic vector text vector encoding on the training tampered dataset to obtain the tampered training document semantic vector of the training tampered dataset. Cross-document feature cross-matching and fusion processing is performed on the tampered training document semantic vector to obtain the tampered intermediate feature representation corresponding to the tampered training document semantic vector. The tampered intermediate feature representation belongs to the intermediate feature representation.

[0091] By using a processing unit that is adapted to the training target for detecting abnormal fields in the document, abnormal field location processing is performed on the tampered intermediate feature representation to obtain the abnormal field location results corresponding to the positions of the multiple data units.

[0092] Based on the location of the second data unit and the abnormal field location results corresponding to the multiple data unit locations, the error value corresponding to the document abnormal field detection training target is determined.

[0093] In this embodiment of the invention, for example, the server obtains an unlabeled document training sample, such as a data pair containing "purchase order" and "supplier historical delivery record". The server performs heterogeneous data alignment processing on this sample using the initial basic document feature extraction component. Its semantic mapper parses "supplier: Adecco Manufacturing" in the purchase order as the "supplier name" feature, and maps it to "delivery reliability rating: excellent" based on the supplier's past on-time delivery record. At the same time, it parses "payment terms: payment upon delivery in 30 days" in the order as the "payment terms" feature, and maps it to "payment period type: standard". These results together constitute the "first sample standardized document dataset", which contains the semantic decomposition results of the sample document at multiple data unit positions, for example, position 1 is "delivery reliability rating: excellent", position 2 is "payment period type: standard".

[0094] Next, the server selects a second data unit location from these multiple data unit locations according to preset rules, to be tampered with. For example, the server selects location 1 ("Delivery Reliability Rating: Excellent") as the tampering target.

[0095] Then, the server performs field tampering processing. It replaces the content at that position with a pre-defined "outlier field value" that clearly conflicts with the original value. For example, it changes "Excellent" to "Severely Lagging". This "Severely Lagging" is an outlier value that may contradict historical facts and the current context (standard payment period). After the tampering, the server obtains a "training tampered dataset" in which the data at position 1 becomes "Delivery Reliability Rating: Severely Lagging", while the data at other positions remains unchanged.

[0096] Subsequently, the server inputs this training tampered dataset into the initial basic document feature extraction component for forward feature propagation. The embedding layer and cross-document feature fusion subcomponent within the component work sequentially to process this dataset containing contradictory information. The fusion subcomponent attempts to understand the relationships between fields, potentially discovering an unreasonable risk mismatch between a "severely delayed" delivery rating and a "standard" payment period. Ultimately, the component outputs a "tampered intermediate feature representation" that encapsulates this inconsistent semantics.

[0097] Subsequently, the server invokes a processing unit adapted to the training objective for detecting abnormal fields in the document. This unit can be an attention-based mechanism or a multi-head classifier. This processing unit receives the altered intermediate feature representation and analyzes it, calculating an "abnormal probability score" for each data unit location in the original standardized dataset. For example, it might output an abnormal probability of 0.92 for location 1 and 0.05 for location 2. The server then converts these probability scores into specific "abnormal field location results," such as determining location 1 as an abnormal field.

[0098] Finally, the server calculates the error value to guide model optimization. It compares the "abnormal field location result" predicted by the processing unit (i.e., position 1 is marked as abnormal) with the actual tampering location (i.e., the second data unit location, which is position 1). Since this is a location task, the server typically uses a binary cross-entropy loss function to calculate the difference between the model's predicted anomalous probability for all locations (including tampered and untampered locations) and the true label (1 for tampered locations, 0 for other locations). This difference value is the "error value corresponding to the document abnormal field detection training target". This error, through backpropagation, prompts the model parameters to be updated, thereby enhancing the model's ability to keenly detect data contradictions and potential anomalies in complex contexts.

[0099] In this embodiment of the invention, the at least one self-supervised training objective includes a heterogeneous enhanced contrastive training objective;

[0100] The step of performing forward feature propagation on the unlabeled document training samples through the initial basic document feature extraction component to obtain intermediate feature representations associated with each self-supervised training objective, and determining at least one error value corresponding to the at least one self-supervised training objective based on the intermediate feature representations, can be implemented through the following example.

[0101] Perturbation data is injected into the document text units of the unlabeled document training samples to obtain perturbation document training samples. Heterogeneous data alignment is performed on the perturbation document training samples to obtain the perturbation-standardized document dataset of the perturbation document training samples.

[0102] Heterogeneous data alignment is performed on the unlabeled document training samples to obtain the first sample standardized document dataset of the unlabeled document training samples;

[0103] The initial basic document feature extraction component performs document semantic vector text vector encoding processing on the perturbation standardized document dataset to obtain the perturbation training document semantic vector of the perturbation standardized document dataset. The perturbation training document semantic vector is then subjected to cross-document feature cross-matching and fusion processing to obtain the perturbation intermediate feature representation corresponding to the perturbation training document semantic vector. The perturbation intermediate feature representation belongs to the intermediate feature representation.

[0104] The initial basic document feature extraction component performs document semantic vector text vector encoding on the first sample standardized document dataset to obtain the original training document semantic vector of the first sample standardized document dataset. The original training document semantic vector is then subjected to cross-document feature cross-matching and fusion processing to obtain the original intermediate feature representation corresponding to the original training document semantic vector. The original intermediate feature representation belongs to the intermediate feature representation.

[0105] Based on the perturbation intermediate feature representation and the original intermediate feature representation, the error value corresponding to the heterogeneous enhanced contrast training target is determined.

[0106] In an embodiment of the invention, for example, the server obtains an unlabeled document training sample, such as a data pair containing a "Purchase Order" (PDF scan) and a corresponding "Logistics Receipt" (JPG image uploaded by a mobile phone). The server injects preset perturbation data into the document text units of this sample to simulate common noise in real-world data collection. For example, the server uses an image processing algorithm to simulate slight stains on the PDF scan, making some numbers in "Purchase Quantity: 1000" blurry; simultaneously, it simulates uneven lighting on the JPG image, making the "Signature of Recipient" field difficult to read. After these operations, the server obtains a "perturbed document training sample" that is visually different but whose business semantics remain unchanged.

[0107] Next, the server performs two heterogeneous data alignment operations in parallel. On one hand, it processes the perturbed document training samples. The semantic mapper in its initial basic document feature extraction component needs to overcome noise interference, correctly parse "Purchase Quantity: 1000" from blurry images and map it to "Purchase Scale: Large Batch," and identify valid recipient information from unevenly lit images. The final output is a structured "Perturbed Standardized Document Dataset." On the other hand, it performs the same alignment process on the original unlabeled document training samples (clear version), outputting a structured "First Sample Standardized Document Dataset." Ideally, these two standardized datasets should be completely identical in semantic content.

[0108] Subsequently, the server performs forward feature propagation and comparison. It inputs the perturbation-normalized document dataset into the initial basic document feature extraction component. The embedding layer within the component converts this dataset into a numerical vector, namely the "perturbation training document semantic vector". This vector is then fed into the cross-document feature fusion sub-component, which integrates the information from the order and the receipt, outputting a high-dimensional "perturbation intermediate feature representation". Simultaneously, the server inputs the first sample normalized document dataset into the same component, undergoes the exact same processing flow, and obtains the "original training document semantic vector" and the corresponding "original intermediate feature representation".

[0109] Finally, the server calculates the contrast error based on these two intermediate feature representations. It uses a contrastive loss function (e.g., InfoNCELoss) for this calculation. The core idea of ​​this function is that the representations of two different variants (original and perturbed versions) of the same document in the feature space should be very close (positive sample pairs), while the feature representations of other randomly selected documents (negative samples) should be significantly far apart. The server treats the "original intermediate feature representation" and the "perturbed intermediate feature representation" as a positive sample pair, while randomly selecting feature representations of other documents from the current training batch as negative samples. The loss function calculates the similarity between positive sample pairs and encourages them to be much higher than the similarity with all negative samples. The calculated difference value is the "error value corresponding to the heterogeneous enhancement contrast training objective." Through backpropagation optimization, the model is trained to ignore surface interference such as formatting and noise, penetratingly capturing stable and consistent deep business semantic relationships across documents.

[0110] In this embodiment of the invention, the at least one self-supervised training objective includes a model distillation-guided training objective;

[0111] The step of performing forward feature propagation on the unlabeled document training samples through the initial basic document feature extraction component to obtain intermediate feature representations associated with each self-supervised training objective, and determining at least one error value corresponding to the at least one self-supervised training objective based on the intermediate feature representations, can be implemented through the following example.

[0112] Heterogeneous data alignment is performed on the unlabeled document training samples to obtain the first sample standardized document dataset of the unlabeled document training samples;

[0113] The initial basic document feature extraction component performs document semantic vector text vector encoding on the first sample standardized document dataset to obtain the original training document semantic vector of the first sample standardized document dataset. The original training document semantic vector is then subjected to cross-document feature cross-matching and fusion processing to obtain the original intermediate feature representation corresponding to the original training document semantic vector. The original intermediate feature representation belongs to the intermediate feature representation.

[0114] Based on the processing unit adapted to the training objective guided by the model distillation, the original intermediate feature representation is subjected to a hazard assessment process to obtain the training hazard assessment result corresponding to the unlabeled document training sample;

[0115] Obtain a teacher model that is compatible with the initial basic document feature extraction component, and perform a hazard assessment on the unlabeled document training samples based on the teacher model to obtain the simulated hazard association index corresponding to the unlabeled document training samples;

[0116] Based on the training hazard assessment results and the simulated hazard correlation index, the error value corresponding to the model distillation-guided training objective is determined.

[0117] In this embodiment of the invention, for example, the server obtains an unlabeled document training sample, such as a dataset containing a "purchase order" for a certain automotive parts, "supplier historical quality inspection reports," and "transit temperature records." The server performs heterogeneous data alignment processing on this sample using an initial basic document feature extraction component (student model). Its semantic mapper parses "material: lithium-ion battery cell" in the purchase order as a "material type" feature, mapping the result to "material risk level: high"; it parses "average batch defect rate: 0.05%" in the quality inspection report as a "quality indicator" feature, mapping the result to "quality performance: excellent"; and it parses "average temperature throughout: 25℃" in the logistics records as a "transportation environment" feature, mapping the result to "temperature control: meets requirements." These results together constitute a structured "first sample standardized document dataset."

[0118] Next, the server performs forward feature propagation on the dataset using the student model. The embedding layer within the student model converts the dataset into a "raw training document semantic vector". This vector is then fed into a cross-document feature fusion subcomponent, which performs a comprehensive correlation analysis on information such as the procurement, quality history, and transportation conditions of the battery cells, ultimately outputting a "raw intermediate feature representation" that can characterize the overall supply chain status of this batch of battery cells.

[0119] Then, the server performs a risk assessment on the original intermediate feature representation based on a processing unit adapted to the model's distillation-guided training objective. This processing unit is typically a lightweight fully connected network layer that receives the intermediate feature representation and outputs a preliminary "training risk assessment result." For example, it might output a scalar value of 0.72, indicating the degree of risk that the student model currently considers to be present in this batch of battery cells.

[0120] Simultaneously, the server acquires a teacher model adapted to the initial basic document feature extraction component in parallel. This teacher model is a powerful model with more parameters, a more complex structure, and has been fully trained on massive amounts of labeled data. The server inputs the same unlabeled document training sample into the teacher model. The teacher model performs a more in-depth and complex analysis process, considering not only the above features but also potentially incorporating external knowledge bases (such as industry recall records) for reasoning, ultimately outputting a more accurate and reliable "simulated hazard correlation index," such as a more authoritative assessment value of 0.85.

[0121] Finally, the server calculates the distillation error to optimize the student model. It compares the training hazard assessment result (0.72) output by the student model with the simulated hazard correlation index (0.85) provided by the teacher model. The server uses the mean squared error loss function to calculate the difference between these two values. This difference is the "error value corresponding to the training objective guided by model distillation". This error signal guides the student model (the initial basic document feature extraction component) to adjust its parameters, including the mapping of the embedding layer and the weights of the fusion sub-components, so that the intermediate feature representation it generates, after simple processing, can approximate the "soft objective" (i.e., the simulated index) given by the teacher model as closely as possible. Through such distillation training with a large number of samples, the student model can absorb the rich domain knowledge and complex pattern recognition capabilities contained in the teacher model, thereby obtaining high-quality feature extraction performance without the need for a large amount of labeled data.

[0122] In this embodiment of the invention, the core feature set of the supply chain is obtained by cross-document feature fusion sub-component in the basic document feature extraction component performing cross-document feature cross-matching and fusion processing on the semantic vector of the document; the basic document feature extraction component belongs to the supply chain penetration supervision model used to conduct supply chain penetration supervision assessment on the supply chain supervision entity; the supply chain penetration supervision model also includes a special assessment feature extraction component serially connected after the cross-document feature fusion sub-component.

[0123] The method of predicting potential risks based on the core feature set of the supply chain according to each penetrating regulatory dimension, and obtaining the potential risk correlation index corresponding to each penetrating regulatory dimension, can be implemented through the following example.

[0124] A special assessment feature extraction component is determined from the supply chain penetration supervision model; the special assessment feature extraction component includes the special assessment model corresponding to each penetration supervision dimension;

[0125] By using specialized assessment models corresponding to each penetrating regulatory dimension, potential risks are predicted for the core feature set of the supply chain, and the potential risk correlation index corresponding to each penetrating regulatory dimension is obtained; a specialized assessment model is used to determine the potential risk correlation index corresponding to a penetrating regulatory dimension.

[0126] In this embodiment of the invention, for example, after obtaining the core feature set of the supply chain through the cross-document feature fusion sub-component, the server then determines the specialized assessment feature extraction component located downstream of the fusion sub-component from the loaded supply chain penetration supervision model. This component is an integrated model module that encapsulates multiple independently trained neural network sub-models, each specifically responsible for a particular penetration supervision dimension. For example, for the supply chain of an electronics manufacturer, this component may integrate four specialized assessment models, corresponding to the four dimensions of "supplier performance risk," "product quality traceability risk," "cash flow compliance risk," and "logistics timeliness and interruption risk," respectively.

[0127] Next, the server uses the core feature set of the supply chain obtained in the previous step as a unified input and distributes it in parallel to each special assessment model within the special assessment feature extraction component. Each model receives the same core feature set, but extracts and focuses on a subset of features relevant to its own regulatory perspective for in-depth computation.

[0128] Specifically, the "Product Quality Traceability Risk" model focuses on core features related to material batches, supplier quality inspection history, and production process correlations. It uses an internal multi-layered neural network to calculate and output a value between 0 and 1, such as 0.76, as a potential risk correlation index for that dimension. This index represents the probability of a break or anomaly in the quality traceability of the entire chain from raw materials to finished products.

[0129] Meanwhile, the "Logistics Timeliness and Disruption Risk" model focuses on analyzing core features related to historical performance of transportation routes, in-transit events, and node congestion. It also performs forward propagation calculations to output a hazard correlation index for this dimension, for example, 0.58.

[0130] The server runs all specialized assessment models in parallel, with each model independently predicting potential hazards for its specific dimension. Finally, the server collects the outputs of all models and aggregates them into a set of hazard correlation indices that correspond one-to-one with the preset regulatory dimensions. For example, the server's results are: Supplier Performance Risk Index 0.65, Product Quality Traceability Risk Index 0.76, Cash Flow Compliance Risk Index 0.42, and Logistics Timeliness and Disruption Risk Index 0.58. These quantified indices provide direct, multi-dimensional risk measurement data for subsequent comprehensive assessments and prioritization.

[0131] In this embodiment of the invention, the following implementation methods are also provided.

[0132] Obtain a second training corpus; the labeled document training samples included in the second training corpus are associated with regulatory dimension target values; the regulatory dimension target values ​​include at least one regulatory dimension label value corresponding to the at least one penetrating regulatory dimension, one penetrating regulatory dimension corresponds to one regulatory dimension label value, and the at least one regulatory dimension label value is determined based on the penetrating regulatory dimension to which the labeled document training sample belongs;

[0133] Obtain a trained basic document feature extraction component, and use the basic document feature extraction component to perform text vector encoding processing on the labeled document training sample to obtain the labeled training document semantic vector of the labeled document training sample. Perform cross-document feature cross-matching and fusion processing on the labeled training document semantic vector to obtain the labeled intermediate feature representation corresponding to the labeled training document semantic vector.

[0134] Obtain at least one initial special assessment model corresponding to the at least one penetrating regulatory dimension, and perform regulatory dimension discrimination on the labeled intermediate feature representation based on the at least one, to obtain the at least one regulatory dimension discrimination confidence level corresponding to the at least one initial special assessment model; one penetrating regulatory dimension corresponds to one initial special assessment model, and one initial special assessment model is used to determine the regulatory dimension discrimination confidence level;

[0135] A second target error value is determined based on the confidence level of the at least one regulatory dimension and the labeled value of the at least one regulatory dimension. Based on the second target error value, the model parameters of the at least one initial special assessment model are optimized, and the optimized at least one initial special assessment model is determined as the at least one special assessment model. The special assessment feature extraction component is determined based on the at least one special assessment model.

[0136] In this embodiment of the invention, for example, the server acquires a second training corpus. The server retrieves a batch of supply chain case data from a historical archive that has been reviewed and manually annotated by domain experts, forming a set of "annotated document training samples." For example, a sample may contain a complete chain of documents for a known delivery delay, including a purchase contract, production plan, logistics authorization letter, and the final customer complaint form. Based on the facts, experts assign specific "regulatory dimension annotation values" to each penetrating regulatory dimension of the sample. For example, the "logistics fulfillment risk" dimension is annotated as 0.95 (high risk), the "product quality risk" dimension as 0.10 (low risk), and the "financial compliance risk" dimension as 0.30. These annotation values ​​are the "regulatory dimension target values" that the model needs to learn.

[0137] Next, the server loads the previously trained basic document feature extraction component with frozen parameters. The server uses this component to process each labeled document training sample: first, it performs heterogeneous data alignment and text vector encoding to generate a "labeled training document semantic vector"; then, through its cross-document feature fusion sub-component, it performs deep feature fusion, outputting a highly refined "labeled intermediate feature representation." This representation carries the comprehensive semantic information of that sample case.

[0138] Then, the server initializes the "initial special assessment model" corresponding to each penetrating regulatory dimension. For example, it creates a small neural network classifier with the same structure but independent parameters for each of the four dimensions: "logistics performance risk," "product quality risk," "financial compliance risk," and "supply chain resilience risk." The server inputs the labeled intermediate feature representations obtained in the previous step into these four initial models. Each model performs forward computation independently and outputs a "regulatory dimension discrimination confidence score" representing its risk assessment for that dimension. For example, the logistics risk model might output 0.88, and the product quality risk model might output 0.15.

[0139] The server then calculates a second objective error value to guide optimization. It compares the decision confidence score of each initial model output with the corresponding actual regulatory dimension's labeled value, dimension by dimension. For example, it calculates the difference between the predicted value of 0.88 for the logistics risk model and the labeled value of 0.95, and the difference between the predicted value of 0.15 for the product quality risk model and the labeled value of 0.10. The server uses a mean squared error loss function to weightedly sum the prediction errors across all dimensions, obtaining an overall second objective error value.

[0140] Finally, based on this second objective error value, the server uses a backpropagation algorithm to update and optimize only the parameters of each initial special assessment model, while keeping the parameters of the basic document feature extraction component unchanged. The server repeats this process using a large number of labeled samples. After multiple rounds of iterative training, the prediction accuracy of each initial special assessment model is significantly improved. The server determines this optimized and converged set of models as the final usable "special assessment models," which together constitute the complete "special assessment feature extraction component." This component is integrated into the supply chain penetration supervision model for accurate, multi-dimensional hazard prediction of the core feature set of the supply chain in the production environment.

[0141] In this embodiment of the invention, the document semantic vector is obtained by performing document semantic vector text vector encoding on the standardized document dataset through the embedding layer in the basic document feature extraction component; the standardized document dataset is obtained by performing heterogeneous data alignment on the end-to-end document cluster through the semantic mapper in the basic document feature extraction component; the basic document feature extraction component belongs to a supply chain penetration supervision model used for supply chain penetration supervision assessment of the supply chain supervision entity; the supply chain penetration supervision model also includes a weight allocation feature extraction component serially connected after the embedding layer;

[0142] The step of assigning weighted anchors to the document semantic vector based on the regulatory dimension driven by the document semantic, and obtaining the regulatory priority coefficients corresponding to each penetrating regulatory dimension, can be implemented through the following example.

[0143] The weight allocation feature extraction component is determined from the aforementioned supply chain penetration supervision model;

[0144] The document semantic vector is weighted and anchored by the document semantic driving regulatory dimension through the weight allocation feature extraction component to obtain the regulatory priority coefficient corresponding to each penetrating regulatory dimension.

[0145] In an embodiment of the invention, for example, after generating a "document semantic vector" representing the semantics of the entire business chain through the embedding layer of the basic document feature extraction component, the server then determines the "weight allocation feature extraction component" serially connected downstream of the embedding layer from the loaded supply chain penetration supervision model. This component is an independent neural network module designed to analyze the overall semantic context of the current document flow and dynamically determine the relative urgency and importance of each supervision dimension.

[0146] The server inputs the generated document semantic vectors into the weighted feature extraction component. This component analyzes the data using its internal multi-layered attention mechanism. For example, for a batch of supply chain documents for new energy vehicle batteries, the component parses the document semantic vectors and identifies several key semantic signals: the purchase order shows an "urgent" label and the quantity is significantly higher than average; the production work order shows that the batch of materials was assigned to a high-priority assembly line; and simultaneously, the associated logistics forecast data indicates that some transportation routes are under a red rainstorm warning.

[0147] Based on a deep understanding of these semantic signals, the weighted feature extraction component performs weighted anchoring calculations. It assesses the potential impact of these business contexts on different regulatory dimensions: "Supply chain resilience risk" (the ability to cope with disruptions) is critical due to the involvement of expedited procurement and critical production lines; "logistics fulfillment risk" has a significantly increased priority for real-time monitoring due to the existence of clear logistics weather warnings; and "product quality risk" is relatively less urgent at present because the batch of materials has complete historical quality inspection reports from the supplier; while "financial compliance risk" is at a normal level of concern because the trading party is a core supplier with long-term cooperation.

[0148] After nonlinear calculations by the internal neural network, the weight allocation feature extraction component outputs a set of normalized "regulatory priority coefficients." For example, it outputs: a logistics fulfillment risk coefficient of 0.40, a supply chain resilience risk coefficient of 0.35, a product quality risk coefficient of 0.15, and a financial compliance risk coefficient of 0.10. The sum of these coefficients is 1, precisely quantifying how regulatory resources and attention should be intelligently allocated across different risk dimensions in the current specific business scenario. This result will be combined with the hazard correlation index for each dimension to calculate the final comprehensive assessment value.

[0149] In this embodiment of the invention, the supply chain penetration supervision model further includes a special assessment feature extraction component for determining the hidden danger correlation index corresponding to each penetration supervision dimension;

[0150] The present invention also provides the following embodiments.

[0151] Obtain a third training corpus; the comprehensive evaluation training samples included in the third training corpus are associated with regulatory target values;

[0152] Obtain a trained basic document feature extraction component, and use the basic document feature extraction component to perform text vector encoding processing on the comprehensive evaluation training sample to obtain the comprehensive training document semantic vector of the comprehensive evaluation training sample. Perform cross-document feature cross-matching and fusion processing on the comprehensive training document semantic vector to obtain the comprehensive intermediate feature representation corresponding to the comprehensive training document semantic vector.

[0153] A trained special assessment feature extraction component is obtained, and at least one special assessment model in the special assessment feature extraction component is used to predict potential hazards in the comprehensive intermediate feature representation to obtain the training hazard correlation index corresponding to each penetrating supervision dimension; a special assessment model is used to determine the training hazard correlation index corresponding to a penetrating supervision dimension.

[0154] The integrated training document semantic vector is weighted and anchored by the document semantic driving regulatory dimension through the initial weight allocation feature extraction component, so as to obtain the training regulatory priority coefficients corresponding to each penetrating regulatory dimension.

[0155] Based on the training hazard correlation index corresponding to each penetrating supervision dimension and the training supervision priority coefficient corresponding to each penetrating supervision dimension, the training comprehensive evaluation value of the sample supply chain supervision subject associated with the comprehensive evaluation training sample is determined.

[0156] A third target error value is determined based on the regulatory target value and the training comprehensive evaluation value. Based on the third target error value, the model parameters of the initial weight allocation feature extraction component are optimized, and the optimized initial weight allocation feature extraction component is determined as the weight allocation feature extraction component. Based on the basic document feature extraction component, the special evaluation feature extraction component, and the weight allocation feature extraction component, the supply chain penetration supervision model is determined.

[0157] In this embodiment of the invention, for example, the server acquires a third training corpus. The server retrieves a batch of more comprehensive "comprehensive assessment training samples" from a historical regulatory case library. Each sample not only contains a complete cluster of supply chain documents, but is also associated with a global "regulatory target value" given by domain experts based on the event's ultimate impact. For example, a sample may record a complete event document flow of a production line shutdown caused by a shortage of critical chips. Based on the severity, scope, and recovery cost of this disruption, experts assign an overall risk level score of 0.90 to this case as a regulatory target value.

[0158] Next, the server loads the previously trained and parameter-frozen basic document feature extraction component and special evaluation feature extraction component. The server uses the basic component to process the comprehensive evaluation training samples, and after heterogeneous alignment, vector encoding and feature fusion, outputs a highly condensed "comprehensive intermediate feature representation".

[0159] The server then inputs this comprehensive intermediate feature representation into a pre-trained specialized assessment feature extraction component. Various specialized assessment models within this component (such as the "supplier risk" and "logistics risk" models) work in parallel, outputting the "training hazard correlation index" for the sample under each penetrating regulatory dimension. For example, the output results are: supplier risk index 0.82, logistics risk index 0.45, and quality risk index 0.20.

[0160] Simultaneously, the server inputs the "comprehensive training document semantic vector" generated by the basic components into the "initial weight allocation feature extraction component" to be trained. This initial component, based on its current understanding of document semantics (e.g., recognizing frequent occurrences of keywords such as "chip," "emergency sourcing," and "production line awaiting materials" in the document flow), calculates and outputs a set of "training supervision priority coefficients." For example, it might initially determine that the supplier risk dimension is the most critical, assigning it a coefficient of 0.50, the logistics risk coefficient of 0.30, and the quality risk coefficient of 0.20.

[0161] Subsequently, the server performs a comprehensive evaluation calculation and error comparison. Using the same logic as the production environment, it weights and sums the training hazard correlation index (0.82, 0.45, 0.20) and the training supervision priority coefficient (0.50, 0.30, 0.20) obtained in the previous step, calculating the "training comprehensive evaluation value" for this sample as 0.82 * 0.50 + 0.45 * 0.30 + 0.20 * 0.20 = 0.605. The server compares this calculated value of 0.605 with the supervision target value of 0.90 marked by experts, calculates the difference between the two (e.g., using mean squared error), and obtains the "third target error value".

[0162] Finally, based on this third objective error value, the server uses a backpropagation algorithm to update and optimize only the model parameters of the initial weight allocation feature extraction component, while keeping the parameters of the basic component and the special evaluation component fixed. Through iterative training with numerous comprehensive cases, the weight allocation component learns how to more accurately adjust the weights of each risk dimension dynamically based on the overall semantics of the document flow, ensuring that the final weighted comprehensive evaluation value continuously approximates the global assessment given by experts. After training, the server designates this optimized component as the final "weight allocation feature extraction component." At this point, the server integrates the trained basic document feature extraction component, special evaluation feature extraction component, and weight allocation feature extraction component according to a pre-defined architecture, forming a complete, production-ready "supply chain penetration monitoring model."

[0163] In this embodiment of the invention, the weight allocation feature extraction component includes an attention mechanism and a feature saliency modeling mechanism;

[0164] The step of assigning weights to the document semantic vector through the weighted feature extraction component to anchor the regulatory dimension driven by the document semantics, and obtaining the regulatory priority coefficients corresponding to each penetrating regulatory dimension, can be implemented through the following example.

[0165] The attention mechanism is used to perform cross-document feature alignment and weighted fusion processing on the semantic vector of the document to obtain the attention matching feature representation corresponding to the semantic vector of the document.

[0166] The attention-matching feature representation is evaluated for saliency using the feature saliency modeling mechanism to obtain the regulatory priority coefficients corresponding to each penetrating regulatory dimension.

[0167] In this embodiment of the invention, for example, the server processes the input document semantic vector through an attention mechanism within the component. This mechanism scans the entire vector, identifying and focusing on key feature fragments most relevant to the current business context. For instance, when processing the semantic vector of a batch of new energy vehicle battery purchase orders, the attention mechanism detects vector fragments representing "purchase order attribute: expedited," "production work order associated production line: flagship model dedicated line," and "logistics forecast information: typhoon warning in the transit area" as highly informative. It highlights these key fragments from the overall vector and performs cross-document semantic alignment and weighted fusion, ultimately generating an "attention-matched feature representation" focused on the current urgent business situation.

[0168] Subsequently, the server performs a deep evaluation of the attention-matched feature representations using a feature saliency modeling mechanism within this component. This mechanism is essentially a multi-task classifier that analyzes the "saliency" or degree of influence of attention-weighted feature combinations on different regulatory dimensions. For example, it determines that the combination of "expedited procurement" and "dedicated production line" significantly increases the attention given to the "supply chain resilience risk" dimension; while the feature "typhoon warning" significantly increases the urgency of the "logistics fulfillment risk" dimension. Simultaneously, it assesses that the contribution of other features (such as the supplier's long-term stable cooperation record) to the "financial compliance risk" dimension is relatively low in the current context. After calculation, the mechanism outputs a set of normalized coefficients, such as: logistics fulfillment risk coefficient 0.40, supply chain resilience risk coefficient 0.35, product quality risk coefficient 0.15, and financial compliance risk coefficient 0.10. These coefficients are the final determined "regulatory priority coefficients corresponding to each penetrating regulatory dimension."

[0169] In this embodiment of the invention, the at least one penetrating regulatory dimension includes a target penetrating regulatory dimension;

[0170] The present invention also provides the following embodiments.

[0171] Obtain a supply chain penetration supervision model for conducting supply chain penetration supervision assessment on supply chain supervision entities, and obtain an initial lightweight assessment network for transferring model knowledge from the supply chain penetration supervision model;

[0172] Obtain a fourth training corpus that corresponds to the target penetration supervision dimension; the fourth training corpus includes dimension-adapted document training samples;

[0173] The supply chain penetration supervision model is used to perform forward inference calculations on the dimension-adapted document training samples to obtain the comprehensive evaluation value corresponding to the dimension-adapted document training samples. The comprehensive evaluation value corresponding to the dimension-adapted document training samples is determined as the synthetic supervision target value used to train the initial lightweight evaluation network.

[0174] The initial lightweight evaluation network is used to perform forward inference calculations on the dimension-adapted document training samples to obtain the comprehensive transfer evaluation value corresponding to the dimension-adapted document training samples.

[0175] Based on the migration comprehensive evaluation value and the synthetic regulatory target value, the model parameters of the initial lightweight evaluation network are optimized, and the optimized initial lightweight evaluation network is determined as the target lightweight evaluation network of the supply chain penetration regulatory model; the target lightweight evaluation network is used to perform supply chain penetration regulatory evaluation on the entire link document cluster under the target penetration regulatory dimension.

[0176] In this embodiment of the invention, for example, the server acquires a fully trained supply chain penetration regulatory model for comprehensive evaluation. Simultaneously, the server initializes an "initial lightweight evaluation network" with a simpler structure and fewer parameters. This network may contain only a basic feature extraction layer and a regression output layer, with computational overhead far less than the complete regulatory model.

[0177] Next, the server determines the target penetration regulatory dimension, such as "logistics fulfillment risk". For this dimension, the server obtains a fourth training corpus, namely a batch of "dimensional-adapted document training samples". These samples are specially selected or generated document clusters that are highly related to logistics fulfillment, such as a large dataset containing logistics documents such as transportation authorization letters, in-transit tracking records, port customs declarations, and customer receipts. These include both normal cases and known delays or abnormal cases.

[0178] Then, the server performs knowledge distillation. It invokes a massive supply chain penetration regulatory model to perform complete forward inference calculations on this batch of dimension-adapted document training samples. For each sample, the complete model undergoes a series of complex calculations, including heterogeneous data alignment, semantic encoding, cross-document fusion, multi-dimensional special assessment, and weight allocation, ultimately outputting a "comprehensive assessment value." The server uses this value as the authoritative "synthetic regulatory target value." For example, for a document sample showing "shipment delay and backup land transportation not activated," the complete model might output a comprehensive assessment value of 0.83.

[0179] Meanwhile, the server inputs the same dimension-adapted document training samples into the initial lightweight evaluation network. This network directly performs rapid feature extraction and calculation on the raw or simply preprocessed document data, outputting a "transfer comprehensive evaluation value". In the early stages of training, this value may differ significantly from the output of the full model, for example, outputting only 0.45.

[0180] Subsequently, the server calculates the optimization loss. It uses loss functions such as mean squared error to calculate the difference between the overall transfer evaluation value (0.45) output by the lightweight network and the synthetic regulatory target value (0.83) provided by the full model. Based on the gradient calculated from this difference, the server uses the backpropagation algorithm to optimize and update all model parameters of the initial lightweight evaluation network.

[0181] The server repeatedly performs the above process using a massive number of dimension-adapted document training samples. After thorough training, the lightweight network learns to mimic the comprehensive judgment logic of the full model on the specific dimension of "logistics fulfillment risk." Finally, the server determines the optimized and converged network as the "target lightweight evaluation network." This network is independently packaged and deployed, specifically for rapid, low-resource-consumption supply chain penetration and regulatory assessment of the entire document cluster involving logistics links, without having to start the entire massive model system.

[0182] In this embodiment of the invention, the dimension-adapted document training sample includes dimension document data instances generated under the target penetration supervision dimension;

[0183] The acquisition of the fourth training corpus, which corresponds to the target penetration regulatory dimension, can be implemented through the following example.

[0184] Obtain the dimension document data instance under the target penetration supervision dimension;

[0185] Candidate document data are determined from the set of document training samples used to perform instance screening;

[0186] The semantic consistency of the candidate document data and the dimensional document data instance is compared by using a sample augmentation model to obtain a consistency metric between the dimensional document data instance and the candidate document data.

[0187] If the consistency metric reaches the data consistency threshold, the candidate document data will be determined as the expanded document data that is compatible with the target penetrating regulatory dimension.

[0188] The expanded document data and the dimensional document data instance are determined as dimensional adaptation document training samples that correspond to and are adapted to the target penetrating supervision dimension.

[0189] In this embodiment of the invention, for example, the server first identifies the target penetration regulatory dimension as "logistics performance risk". Subsequently, the server retrieves typical "dimensional document data instances" that have been confirmed by experts under this dimension from the historical anomaly event database. For example, a complete set of documents recording severe shipping delays caused by a typhoon last year, which ultimately led to customer claims, including the shipping company's notice of delay, the land transport authorization letter for emergency diversion, and the compensation agreement.

[0190] Next, the server filters out a large number of "alternative document data" that are close in time and related to business from a massive general historical document database (i.e., the "document training sample set"). This alternative data may include normal logistics waybills, customs declaration records, etc., and is not marked as belonging to risk cases.

[0191] Then, the server invokes a pre-trained "sample augmentation model." This model is a deep semantic matching network. The server simultaneously inputs the dimensional instance (the typhoon delay case) and each candidate document into the model. The model performs deep semantic encoding and comparison on both, analyzing their similarity in business entities (such as port of origin, port of destination, carrier type) and event patterns (such as mode of transport connections), and outputs a "consistency metric," which is a score between 0 and 1.

[0192] The server presets a "data consistency threshold", such as 0.75. When the semantic consistency metric of a candidate document data (such as a waybill showing "goods are delayed due to port strike") and the typhoon delay instance reaches 0.80, the server determines that the candidate data is highly compatible with the target dimension in terms of risk mode, and thus identifies it as "expanded document data".

[0193] Finally, the server merges the original dimensional document data instances with all the selected expanded document data to form a set of "dimensional-adapted document training samples" used to train the lightweight assessment network dedicated to "logistics fulfillment risk".

[0194] In this embodiment of the invention, the determination of the comprehensive evaluation value of the supply chain supervision entity based on the hidden danger correlation index and the supervision priority coefficient corresponding to each penetrating supervision dimension can be implemented through the following example.

[0195] The weighted evaluation value for each penetrating regulatory dimension is obtained by performing weighted fusion on the hidden danger correlation index corresponding to each penetrating regulatory dimension through the regulatory priority coefficient corresponding to each penetrating regulatory dimension; the weight feature value corresponding to a penetrating regulatory dimension is used to perform weighted fusion on the hidden danger correlation index corresponding to the corresponding penetrating regulatory dimension.

[0196] Based on the weighted evaluation values ​​corresponding to each of the aforementioned penetrating regulatory dimensions, the comprehensive evaluation value of the supply chain regulatory entity is determined.

[0197] In this embodiment of the invention, for example, the server has obtained the potential risk correlation index corresponding to each penetrating regulatory dimension, such as: supplier performance risk index 0.65, product quality risk index 0.76, logistics performance risk index 0.60, and supply chain resilience risk index 0.70. Simultaneously, the server has also obtained the corresponding regulatory priority coefficients through a weighted feature extraction component, such as 0.15, 0.25, 0.40, and 0.20 respectively.

[0198] First, the server performs weighted fusion for each dimension. It multiplies the risk correlation index of each dimension with its corresponding regulatory priority coefficient to obtain the "weighted assessment value" for that dimension. The specific calculations are as follows: the weighted assessment value for supplier performance risk is 0.65 * 0.15 = 0.0975; the weighted assessment value for product quality risk is 0.76 * 0.25 = 0.19; the weighted assessment value for logistics performance risk is 0.60 * 0.40 = 0.24; and the weighted assessment value for supply chain resilience risk is 0.70 * 0.20 = 0.14.

[0199] The server then sums the weighted evaluation values ​​for all dimensions to determine the final "comprehensive evaluation value." The calculation process is: 0.0975 + 0.19 + 0.24 + 0.14 = 0.6675. This value of 0.6675 represents the quantitative result of the overall risk status of the supply chain regulatory entity. This comprehensive evaluation value is then input into the regulatory decision output module to trigger the corresponding alarm or report generation process.

[0200] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned supply chain penetration monitoring method based on a document engine. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0201] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A supply chain penetration monitoring method based on a document engine, characterized in that, include: Deploy and launch a document engine for supply chain penetration supervision, which integrates a full-link document interaction module, a supply chain penetration supervision model, and a supervision decision output module; The end-to-end document interaction module connects to the end-to-end data collection node of the supply chain supervision entity to obtain the end-to-end document cluster of the supply chain supervision entity; Text vector encoding is performed on the entire document cluster of the supply chain supervision entity to obtain the document semantic vector corresponding to the entire document cluster; Perform cross-document feature cross-matching and fusion processing on the semantic vector of the document to obtain the core feature set of the supply chain corresponding to the semantic vector of the document. Determine at least one penetrating supervision dimension that corresponds to and is compatible with the full-link document cluster. Based on each penetrating supervision dimension, perform hazard prediction on the core feature set of the supply chain to obtain the hazard correlation index corresponding to each penetrating supervision dimension. The document semantic vector is weighted and anchored according to the document semantic driving regulatory dimension to obtain the regulatory priority coefficient corresponding to each penetrating regulatory dimension; Based on the hidden danger correlation index and the regulatory priority coefficient corresponding to each of the penetrating regulatory dimensions, the comprehensive evaluation value of the supply chain regulatory entity is determined. The comprehensive assessment value is input into the regulatory decision output module of the document engine, which generates corresponding supply chain penetration supervision alarm information, compliance assessment report or rectification suggestion list, and feeds it back to the terminal device of the supply chain supervision entity through the end-to-end document interaction module.

2. The method according to claim 1, characterized in that, The text vector encoding process performed on the entire document cluster of the supply chain supervision entity to obtain the document semantic vector corresponding to the entire document cluster includes: Heterogeneous data alignment is performed on the full-link document cluster to obtain a standardized document dataset of the full-link document cluster; The standardized document dataset is subjected to document semantic vector text vector encoding processing to obtain the document semantic vector of the standardized document dataset.

3. The method according to claim 2, characterized in that, The step of aligning heterogeneous data of the end-to-end document cluster to obtain a standardized document dataset for the end-to-end document cluster includes: When the full-chain document cluster of the supply chain regulatory entity is obtained, a supply chain penetration regulatory model for conducting supply chain penetration regulatory assessment of the supply chain regulatory entity is obtained; the supply chain penetration regulatory model includes a basic document feature extraction component. The basic document feature extraction component determines the document attribute category that corresponds to and is suitable for the full-link document cluster, and extracts the document text unit of the full-link document cluster based on the document attribute category; The document text unit is semantically decomposed using the basic document feature extraction component to obtain the semantic decomposition result corresponding to the document text unit. The basic document feature extraction component performs structured expression processing on the semantic decomposition results of the document text units to obtain the standardized document dataset of the full-link document cluster.

4. The method according to claim 3, characterized in that, The basic document feature extraction component includes a semantic mapper; the document text unit includes at least one document text unit, and the at least one document text unit includes a target document text unit. The step of performing semantic decomposition on the document text unit using the basic document feature extraction component to obtain the semantic decomposition result corresponding to the document text unit includes: Based on the document attribute category to which the target document text unit belongs, the semantic feature domain corresponding to the target document text unit is determined; the semantic feature domain includes multiple feature values, and each feature value is associated with a corresponding attribute value domain partition. The semantic mapper determines the attribute value domain partition to which the target document text unit belongs, and the feature value corresponding to the attribute value domain partition to which the target document text unit belongs is determined as the document semantic decomposition result corresponding to the target document text unit.

5. The method according to claim 1, characterized in that, The core feature set of the supply chain is obtained by cross-document feature fusion sub-component in the basic document feature extraction component, which performs cross-document feature cross-matching and fusion processing on the semantic vector of the document; the basic document feature extraction component belongs to the supply chain penetration supervision model used to conduct supply chain penetration supervision assessment on the supply chain supervision entity; The method further includes: Obtain the first training corpus; the unlabeled document training samples included in the first training corpus are document training samples without labeled regulatory target values; Obtain at least one self-supervised training objective for the feature extraction component of the initial basic document; The initial basic document feature extraction component performs forward feature propagation on the unlabeled document training samples to obtain intermediate feature representations associated with each self-supervised training objective. Based on the intermediate feature representations, at least one error value corresponding to the at least one self-supervised training objective is determined; one self-supervised training objective corresponds to one error value. A first target error value is determined based on the at least one error value. The model parameters of the initial basic document feature extraction component are optimized based on the first target error value. The optimized initial basic document feature extraction component is then determined as the basic document feature extraction component.

6. The method according to claim 5, characterized in that, The at least one self-supervised training objective includes a heterogeneous enhanced contrastive training objective; The step of performing forward feature propagation on the unlabeled document training samples through the initial basic document feature extraction component to obtain intermediate feature representations associated with each self-supervised training objective, and determining at least one error value corresponding to the at least one self-supervised training objective based on the intermediate feature representations, includes: Perturbation data is injected into the document text units of the unlabeled document training samples to obtain perturbation document training samples. Heterogeneous data alignment is performed on the perturbation document training samples to obtain the perturbation-standardized document dataset of the perturbation document training samples. Heterogeneous data alignment is performed on the unlabeled document training samples to obtain the first sample standardized document dataset of the unlabeled document training samples; The initial basic document feature extraction component performs document semantic vector text vector encoding processing on the perturbation standardized document dataset to obtain the perturbation training document semantic vector of the perturbation standardized document dataset. The perturbation training document semantic vector is then subjected to cross-document feature cross-matching and fusion processing to obtain the perturbation intermediate feature representation corresponding to the perturbation training document semantic vector. The perturbation intermediate feature representation belongs to the intermediate feature representation. The initial basic document feature extraction component performs document semantic vector text vector encoding on the first sample standardized document dataset to obtain the original training document semantic vector of the first sample standardized document dataset. The original training document semantic vector is then subjected to cross-document feature cross-matching and fusion processing to obtain the original intermediate feature representation corresponding to the original training document semantic vector. The original intermediate feature representation belongs to the intermediate feature representation. Based on the perturbation intermediate feature representation and the original intermediate feature representation, the error value corresponding to the heterogeneous enhanced contrast training target is determined.

7. The method according to claim 5, characterized in that, The at least one self-supervised training objective includes a model distillation-guided training objective; The step of performing forward feature propagation on the unlabeled document training samples through the initial basic document feature extraction component to obtain intermediate feature representations associated with each self-supervised training objective, and determining at least one error value corresponding to the at least one self-supervised training objective based on the intermediate feature representations, includes: Heterogeneous data alignment is performed on the unlabeled document training samples to obtain the first sample standardized document dataset of the unlabeled document training samples; The initial basic document feature extraction component performs document semantic vector text vector encoding on the first sample standardized document dataset to obtain the original training document semantic vector of the first sample standardized document dataset. The original training document semantic vector is then subjected to cross-document feature cross-matching and fusion processing to obtain the original intermediate feature representation corresponding to the original training document semantic vector. The original intermediate feature representation belongs to the intermediate feature representation. Based on the processing unit adapted to the training objective guided by the model distillation, the original intermediate feature representation is subjected to a hazard assessment process to obtain the training hazard assessment result corresponding to the unlabeled document training sample; Obtain a teacher model that is compatible with the initial basic document feature extraction component, and perform a hazard assessment on the unlabeled document training samples based on the teacher model to obtain the simulated hazard association index corresponding to the unlabeled document training samples; Based on the training hazard assessment results and the simulated hazard correlation index, the error value corresponding to the model distillation-guided training objective is determined.

8. The method according to claim 1, characterized in that, The core feature set of the supply chain is obtained by cross-document feature fusion sub-component in the basic document feature extraction component, which performs cross-document feature cross-matching and fusion processing on the semantic vector of the document; the basic document feature extraction component belongs to the supply chain penetration supervision model used to conduct supply chain penetration supervision assessment on the supply chain supervision entity; the supply chain penetration supervision model also includes a special assessment feature extraction component that is serially connected after the cross-document feature fusion sub-component. The method involves predicting potential risks based on each penetrating regulatory dimension of the core feature set of the supply chain, and obtaining the potential risk correlation index corresponding to each penetrating regulatory dimension, including: Determine the special assessment feature extraction components from the aforementioned supply chain penetration supervision model; The special assessment feature extraction component includes the special assessment model corresponding to each penetrating regulatory dimension; By using specialized assessment models corresponding to each penetrating regulatory dimension, potential risks are predicted for the core feature set of the supply chain, and the potential risk correlation index corresponding to each penetrating regulatory dimension is obtained; a specialized assessment model is used to determine the potential risk correlation index corresponding to a penetrating regulatory dimension.

9. The method according to claim 1, characterized in that, The document semantic vector is obtained by encoding the standardized document dataset with document semantic vector text vectors through the embedding layer in the basic document feature extraction component; the standardized document dataset is obtained by aligning the heterogeneous data of the end-to-end document cluster through the semantic mapper in the basic document feature extraction component; the basic document feature extraction component belongs to a supply chain penetration supervision model used for supply chain penetration supervision assessment of the supply chain supervision entity; the supply chain penetration supervision model also includes a weighted feature extraction component serially connected after the embedding layer; The step of assigning regulatory dimension weights and anchoring to the document semantic vector to obtain the regulatory priority coefficients corresponding to each penetrating regulatory dimension includes: The weight allocation feature extraction component is determined from the aforementioned supply chain penetration supervision model; The document semantic vector is weighted and anchored by the document semantic driving regulatory dimension through the weight allocation feature extraction component to obtain the regulatory priority coefficient corresponding to each penetrating regulatory dimension.

10. A server system, characterized in that, The server system includes a computer program, which, when running, controls the server system to execute the method described in any one of claims 1-9.

Images