A human-machine collaborative manufacturing industry document compliance review method and system

Abstract

This invention relates to the field of intelligent manufacturing software technology, and discloses a human-machine collaborative method and system for compliance review of manufacturing documents. The method includes the following steps: first, receiving a compliance review request input by a user, the request including at least a workflow identifier and a manufacturing document to be reviewed; then, performing sensitive information detection on the manufacturing document; after determining that the manufacturing document passes the sensitive information detection, extracting the document content as initial processing data; and finally, reading the corresponding workflow definition file based on the workflow identifier in the compliance review request. By reading the corresponding workflow definition file based on the workflow identifier and dynamically constructing an executable state diagram, the system can quickly adapt to the review scenarios of various heterogeneous manufacturing documents, such as mechanical drawings, without modifying the core code, simply by adding or modifying configuration files, thereby reducing the system's development and maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing software technology, specifically to a human-machine collaborative method and system for compliance review of manufacturing documents. Background Technology

[0002] In modern manufacturing, compliance review of important documents such as R&D drawings, electronic design packages, bills of materials (BOM), and standard operating procedures (SOPs) is a crucial step in ensuring product quality, production safety, and compliance with industry standards. Traditional review models typically rely on technical personnel in different positions to manually compare and verify each item based on their respective business experience. However, faced with multiple versions, multiple standards, and a massive amount of heterogeneous documents, this review model, which is highly dependent on personal experience, not only makes it difficult to maintain consistent review standards, but also easily leads to the risk of omissions in review items and version mismatches during multiple iterations. At the same time, the review process and decision-making basis often cannot be fully recorded for subsequent quality traceability.

[0003] To overcome the efficiency and traceability issues of manual review, some manufacturing companies have tried to introduce digital review information systems. Currently, most review systems generally adopt a hard-coded process implementation method at the underlying architecture. This means that the system logic is deeply bound to specific business scenarios. When companies face new customer specifications or need to add a review process for a certain type of heterogeneous document, they are often required to modify the backend program code from the bottom layer, add dedicated data interfaces, and customize the front-end interactive page. This rigid system architecture results in extremely poor cross-scenario reusability. Not only is the software delivery cycle long, but once the business process is slightly adjusted, the entire system must be re-released, making it difficult to adapt to the business needs of modern manufacturing.

[0004] Meanwhile, with the development of artificial intelligence technology, some cutting-edge solutions have begun to try to apply large language models directly to document review or rely on a single form engine for data flow. However, in manufacturing scenarios with stringent quality audit requirements, the core challenges are not limited to simple document text recognition and extraction. Review methods that rely solely on large language models for question answering are prone to errors due to the randomness and uninterpretability of generated content, making it difficult to guarantee the absolute rule priority principle in industry standards. Traditional form systems are also difficult to effectively parse complex heterogeneous data inputs such as mechanical drawings, deep directory file packages, and dual-source list comparisons. Therefore, this invention designs a human-machine collaborative manufacturing document compliance review method and system to address the aforementioned problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a human-machine collaborative method and system for compliance review of manufacturing documents, which solves the problems of hard-coded processes, fragmented review scenarios, lack of a unified point of contact for manual review, and difficulty in closing the loop on security and auditing in manufacturing document review.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a human-machine collaborative method for compliance review of manufacturing documents, comprising the following steps: S1. First, receive a compliance review request input by the user. The compliance review request includes at least a workflow identifier and a manufacturing document to be reviewed. S2. Subsequently, sensitive information detection is performed on the manufacturing documents. After determining that the manufacturing documents pass the sensitive information detection, the document content of the manufacturing documents is extracted as initial processing data. S3. Based on the workflow identifier in the compliance review request, read the corresponding workflow definition file, and parse the workflow definition file to construct an executable state diagram for performing compliance review. The executable state diagram contains multiple processing nodes. S4. Based on the executable state diagram, the initial processing data is traversed and executed node by node. During the traversal and execution of the processing nodes, preset structured rules are called to automatically determine the initial processing data in order to extract compliance defect evidence in the manufacturing documents and generate automatic review conclusions based on the compliance defect evidence. S5. During the process of traversing and executing the processing nodes, when a preset manual confirmation node in the workflow definition file is triggered, the current automated workflow action is suspended, and the compliance defect evidence and the automatic review conclusion are output and displayed to receive manual confirmation instructions from users based on human-machine collaboration feedback. The workflow definition file supports inserting one or more manual confirmation nodes between any processing nodes. The position and number of manual confirmation nodes can be configured by modifying the metadata of the workflow definition file without modifying the underlying execution code. S6. Upon receiving the manual confirmation instruction, resume the flow of the executable state diagram until the traversal is complete, combine the automatic review conclusion with the manual confirmation instruction to generate the final compliance review deliverable, and write the compliance review request, the compliance review deliverable, and the execution trajectory information of the review process into the anti-tampering audit chain.

[0007] Preferably, the compliance review request received from the user in step S1 specifically includes: Receive the process type selected by the user in the interactive interface and use it as the workflow identifier; Based on the format and quantity of the uploaded manufacturing documents and the preset workflow input constraints, the workflow identifier is validated for reasonableness. Sensitive information detection of the manufacturing documents can only be triggered after the validation matches. The reasonableness validation includes: matching the file format of the uploaded documents with the accepted_formats declared in the input_schema in the workflow definition file; and validating the quantity of uploaded documents against the required field list declared in the input_schema.

[0008] Preferably, the sensitive information detection performed on the manufacturing documents in step S2 specifically includes: A business keyword matching engine and a regular expression engine are used to perform dual recognition and detection on the content of the manufacturing documents. If the blocking rule is met, the review process will be terminated; if the desensitization rule is met, the combination of personal sensitive information and business sensitive words will be replaced while preserving the context structure.

[0009] Preferably, extracting the document content of the manufacturing documents as initial processing data in step S2 specifically includes: The graphic and textual parameter information of the manufacturing documents is extracted using optical character recognition or document parsing methods to generate a structured summary, which is then used as the initial processing data. During subsequent traversal and execution of the processing nodes, the system only inputs the extracted initial processing data into the large language model node, prohibiting the original manufacturing documents from entering the large language model node, so as to achieve secure isolation of the source files.

[0010] Preferably, step S3, which involves parsing the workflow definition file and constructing it into an executable state diagram for performing compliance reviews, specifically includes: The underlying atomic processing nodes are automatically scanned and discovered through a preset node registry. The configuration parameters in the workflow definition file, written in metadata format, are dynamically bound to the corresponding atomic processing nodes to generate the executable state diagram that transmits intermediate results using a shared state dictionary.

[0011] Preferably, the automatic determination of the initial processed data by calling preset structured rules in step S4 specifically includes: The structured rules are executed first to perform deterministic matching on the initial processed data, and to locate and extract the compliance defect evidence with a precise location and rule number; When the structured rules cannot cover the ambiguous semantic scenarios in the initial processed data, the large language model node is triggered by conditional routing to perform semantic supplementary interpretation, and the result is merged with the structured rule judgment result to generate the automatic review conclusion.

[0012] Preferably, receiving the user's manual confirmation instruction based on human-machine collaboration feedback in step S5 specifically includes: The front-end interactive interface highlights the evidence of compliance deficiencies and the corresponding risk locations, and provides a decision interface for users to perform approval or rejection operations; The user's actions on the decision interface are captured as the manual confirmation instruction and updated in the shared state dictionary of the executable state graph to release the suspended state.

[0013] Preferably, step S6, which involves writing the compliance review request, the compliance review deliverables, and the execution trajectory information of the review process into the tamper-proof audit chain, specifically includes: Single-line audit logs are recorded using an append-only method. Each audit log entry contains metadata about the current compliance review request, the action performed, and a hash reference to the previous audit record. The integrity of the entire lifecycle operation record is verified by chaining the hash reference values.

[0014] Preferably, the step of generating the final compliance review deliverable by combining the automatic review conclusions and the manual confirmation instructions is further limited as follows: The compliance review deliverables include at least a user-facing issue rectification list and an internal audit artifact containing the audit logs; Furthermore, by replacing the workflow definition file and structured rules, it supports multi-scenario compliance review of mechanical drawings, electronic design packages, bills of materials, and standard operating procedures, as well as the generation of corresponding compliance review deliverables.

[0015] Preferably, a human-machine collaborative manufacturing document compliance review system includes a processor and a memory. The memory stores computer instructions, which, when executed by the processor, implement the following functional modules: The access request module is used to receive compliance review requests input by users and obtain workflow identifiers and manufacturing documents to be reviewed. A security filtering module is used to perform sensitive information detection on the manufacturing documents and extract the document content of the manufacturing documents as initial processing data after the detection is passed. The workflow compilation module is used to read the corresponding workflow definition file based on the workflow identifier in the compliance review request, and parse and construct it into an executable state diagram containing multiple processing nodes; The execution engine module is used to perform node traversal execution on the initial processing data based on the executable state diagram, call preset structured rules to automatically determine and extract evidence of compliance defects, generate automatic review conclusions, and suspend the current automated workflow when a manual confirmation node is triggered. The human-computer interaction module is used to output and display the compliance defect evidence and the automatic review conclusion. After receiving the manual confirmation instruction from the user, it triggers the execution engine module to resume the process until the traversal is completed. The output and audit module is used to comprehensively generate the final compliance review deliverables and write the compliance review request, the compliance review deliverables, and the execution trajectory information of the review process into the tamper-proof audit chain.

[0016] This invention provides a human-machine collaborative method and system for compliance review of manufacturing documents. It offers the following advantages: 1. This invention achieves complete decoupling between business process definition and underlying execution code by reading the corresponding workflow definition file based on the workflow identifier and dynamically constructing an executable state diagram. This allows the system to quickly adapt to the review scenarios of various heterogeneous manufacturing documents such as mechanical drawings, electronic design packages, and BOM lists without modifying the core code, simply by adding or modifying configuration files. This reduces the average time of 3-6 months under the traditional hard-coded model to just a few hours to a few days by writing workflow definition files and structured rule files, thereby reducing the development and maintenance costs of the system.

[0017] 2. This invention establishes a verification system based on chained hash comparison by writing records including compliance review requests, review deliverables, and execution action trajectories into an anti-tampering audit chain using a single-line log with an appended hash reference value. This ensures that no operation throughout the entire review lifecycle can be silently overwritten or tampered with, thereby meeting the manufacturing industry's requirement for a traceable audit loop for the review process.

[0018] 3. This invention prioritizes calling preset structured rules during node traversal to automatically determine and extract precise evidence of compliance defects. Only in cases of ambiguous semantics that cannot be covered by the rules is a large language model used as a fallback. This allows more than 80% of the standardized review to be completed automatically by the rules, effectively avoiding the error risks that are easily generated by pure large language model question-and-answer review. This ensures that the review conclusions have extremely high accuracy, traceability of evidence location, and interpretability of rules.

[0019] 4. This invention constructs a human-machine collaboration mechanism in the workflow definition file by flexibly pre-setting manual confirmation nodes. When the node is triggered, the current automated workflow action is actively suspended, and the automatic review conclusion is output and displayed to receive manual confirmation instructions from the user. This mechanism not only leverages the efficiency of machines in processing complex and standardized data, but also ensures the accuracy of the final high-risk critical decision.

[0020] 5. This invention performs dual detection of sensitive information at the front end of the review process and forces the extraction of document content as initial processing data before node transfer. This ensures that intelligent processing nodes such as large language models can only consume structured summaries and cannot directly access the original manufacturing source files. This effectively prevents the leakage of core design drawings, trade secrets and privacy data in the logical chain and protects the security of the enterprise's digital assets. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system module architecture of the present invention; Figure 3 This is a schematic diagram of the workflow YAML structure of the present invention; Figure 4 This is a schematic diagram illustrating the secure isolation and structured digest transmission of source files in this invention; Figure 5 This is a schematic diagram illustrating the conditional routing and report switching of the present invention; Figure 6 This is a schematic diagram illustrating the relationship between rule priority review and human-machine collaboration in this invention; Figure 7 This is a schematic diagram illustrating the compilation and execution relationship of the present invention; Figure 8 This is a schematic diagram illustrating the coverage relationship of multiple scenario embodiments of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see the appendix Figure 1 - Appendix Figure 8 This invention provides a human-machine collaborative method for compliance review of manufacturing documents, comprising the following steps: S1. First, receive the compliance review request input by the user. The compliance review request includes at least a workflow identifier and the manufacturing documents to be reviewed. This step of receiving the compliance review request input by the user specifically includes: receiving the process type selected by the user in the interactive interface and using it as the workflow identifier; combining the format and quantity of the uploaded manufacturing documents and the preset workflow input constraints, performing a reasonableness check on the workflow identifier. Only after the check matches can the detection of sensitive information in the manufacturing documents be triggered. The reasonableness check includes: matching the file format of the uploaded documents with the accepted_formats declared in the input_schema in the workflow definition file; and checking the quantity of uploaded documents against the required field list declared in the input_schema. Specifically, in various scenarios in the manufacturing industry (such as mechanical drawing review, electronic design package review, BOM list consistency review, and SOP document review), in order to ensure clear responsibilities and audit traceability, this invention does not adopt the "pure AI automatic guessing of intent" mode; The receiving process is divided into two levels of routing collaboration: In the first level, the user explicitly selects the type of review process to be executed on the interactive interface (e.g., selecting "BOM_Review"), and the system uses this as a workflow identifier; In the second level, the system uses an internal agent or validator to perform a validity check on the format of the manufacturing documents uploaded by the user (e.g., Excel, PDF, or CAD exported files), the number of files (e.g., for a dual BOM review, both EBOM and MBOM files must be uploaded simultaneously), and preset input constraints. Only when the input data completely matches the workflow identifier can subsequent processing begin.

[0024] S2. Subsequently, sensitive information detection is performed on the manufacturing documents. After confirming that the manufacturing documents pass the sensitive information detection, the document content of the manufacturing documents is extracted as initial processing data. The sensitive information detection of the manufacturing documents in this step specifically includes: using a business keyword matching engine and a regular expression engine to perform dual recognition and detection on the content of the manufacturing documents; if a blocking rule is hit, the review process is terminated; if a desensitization rule is hit, a replacement process that preserves the context structure is performed on the combination of personal sensitive information and business sensitive words. The extraction of the document content of the manufacturing documents as initial processing data in this step specifically includes: extracting the graphic and textual parameter information of the manufacturing documents through optical character recognition or document parsing to generate a structured summary, which is used as the initial processing data; when traversing and executing processing nodes in subsequent steps, the system only inputs the extracted initial processing data into the large language model node, prohibiting the original manufacturing documents from entering the large language model node, so as to achieve secure isolation of the source files; Specifically, because manufacturing documents contain trade secrets and personal privacy, the system must implement strict security controls before extracting document content. Sensitive information detection employs a dual mechanism of a "business keyword matching engine" and a "regular expression engine." The keyword engine is responsible for identifying sensitive business expressions (such as cost, quotation, profit, unit price, etc.); the regular expression engine is responsible for capturing standardized format data (such as ID card number, bank card number). If extremely high-risk features are detected, the system directly blocks the review. If it is a review task that needs to be approved, a de-identification and replacement algorithm that "preserves the context structure" is adopted (for example, replacing specific monetary data with the [confidential amount] identifier) ​​to prevent the original value from spreading further. After passing security checks, the system uses Optical Character Recognition (OCR) or a document parser to read the source file and convert it into structured summary data (such as JSON format) as initial processing data. Through this mechanism, the system establishes the principle of source file isolation: in subsequent review processes, especially the Large Language Model (LLM) node, it can only receive structured initial processing data and can never directly touch or read the original design drawings or data packets, thus eliminating the risk of drawings being leaked or absorbed by the model from the physical link.

[0025] S3. Based on the workflow identifier in the compliance review request, read the corresponding workflow definition file and parse the workflow definition file to construct an executable state diagram for performing compliance review. The executable state diagram contains multiple processing nodes. This step of parsing the workflow definition file to construct an executable state diagram for performing compliance review specifically includes: automatically scanning and discovering the underlying atomic processing nodes through a preset node registry; dynamically binding the configuration parameters in the workflow definition file written in metadata format to the corresponding atomic processing nodes, and generating an executable state diagram that passes intermediate results through a shared state dictionary. Specifically, to achieve zero-code expansion for review scenarios, this system completely decouples business processes from underlying code. Workflow definition files are written in metadata formats such as YAML. Their top-level structure includes a unique workflow identifier, input validation rule declarations (input_schema), a list of node definitions (nodes), and deliverable definitions (deliverables). Upon system startup, the node registry automatically scans and loads reusable atomic processing nodes (e.g., file loading nodes, structured parsing nodes, rule checking nodes, etc.). When constructing the executable state graph, the workflow compilation module dynamically binds the parameters configured in the YAML file (such as specific review rule paths and threshold settings) to the corresponding atomic processing nodes. Nodes exchange intermediate result data through a shared state dictionary, thereby generating a directed state graph that can be driven by the engine. Depending on the actual implementation, the top-level structure includes at least the following fields: workflow_id, name, description, input_schema, nodes, and deliverables. The explanations of each layer are shown in the table below.

[0026] S4. Based on the executable state diagram, the initial processing data is traversed and executed node by node. During the traversal and execution of processing nodes, preset structured rules are invoked to automatically determine the initial processing data in order to extract compliance defect evidence from manufacturing documents, and an automatic review conclusion is generated based on the compliance defect evidence. The automatic determination of the initial processing data by invoking preset structured rules in this step specifically includes: prioritizing the execution of structured rules to perform deterministic matching on the initial processing data, locating and extracting compliance defect evidence with a precise location and rule number; when the structured rules cannot cover the ambiguous semantic scenarios in the initial processing data, the large language model node is triggered through conditional routing to perform semantic supplementary interpretation, and the result is merged with the structured rule determination result to generate an automatic review conclusion. Specifically, to overcome the uncertainty of the results generated by a pure large language model, this invention proposes a hybrid review mechanism of "rule priority and large model as a backup". During the node traversal execution phase, the structured rule nodes are given priority in the judgment. The structured rules include field non-empty verification, version number consistency comparison, threshold out-of-bounds judgment, etc. When a rule is hit, the system accurately extracts the evidence of compliance defects (including the specific page number / coordinate position of the defect and the rule number violated) and generates a deterministic automatic review conclusion. When structured rules cannot fully cover the ambiguous semantic scenarios of the current document, or when it is necessary to generate natural language rectification suggestions for defects, the system sends the relevant initial processing data and rule hit results to the large language model node for semantic supplementation through a preset conditional routing node (such as report_mode_router), and finally merges them to form an automatic review conclusion with high interpretability.

[0027] S5. During the traversal of execution processing nodes, when a preset manual confirmation node in the workflow definition file is triggered, the current automated workflow action is suspended, and the compliance defect evidence and automatic review conclusion are output and displayed to receive manual confirmation instructions from users based on human-machine collaboration feedback. This step of receiving manual confirmation instructions from users based on human-machine collaboration feedback specifically includes: highlighting the compliance defect evidence and corresponding risk location on the front-end interactive interface, and providing a decision interface for users to perform approval or rejection operations; capturing the user's operation on the decision interface as a manual confirmation instruction, and updating it to the shared state dictionary of the executable state diagram to release the suspended state. The workflow definition file supports inserting one or more manual confirmation nodes between any processing nodes. The position and number of manual confirmation nodes can be configured by modifying the metadata of the workflow definition file without modifying the underlying execution code. Specifically, this system is not a completely automated black box, but rather returns the final decision-making power to reviewers with professional experience. Specific types of "human confirmation nodes" (HUMAN_CONFIRM) can be flexibly inserted into the workflow definition file. When the execution engine reaches this node, it will automatically establish an execution interruption point (suspending the automated workflow). At this time, the human-computer interaction module will highlight the compliance defect evidence and automatic review conclusions extracted in the early stage on the front-end interface. After verification, the professional reviewer will click "pass" or "reject" on the interface. The system captures these user operations as human confirmation instructions and updates them to the shared state dictionary of the executable state diagram, completing the collaborative closed loop of "human-in-the-Loop".

[0028] S6. Upon receiving a manual confirmation instruction, resume the flow of the executable state diagram until traversal is complete. Integrate the automatic review conclusions and the manual confirmation instruction to generate the final compliance review deliverable. Write the compliance review request, compliance review deliverable, and execution trajectory information of the review process into the tamper-proof audit chain. Specifically, this step involves: recording single-line audit logs using an append-only method. Each audit log entry contains the current compliance review request metadata, the execution action, and a hash reference value pointing to the previous audit record. Verify the completeness of the entire lifecycle of the review operation records through chain comparison of the hash reference values. For the step of generating the final compliance review deliverable by integrating the automatic review conclusions and the manual confirmation instruction, further limitations are made as follows: The compliance review deliverable must at least include a user-facing issue rectification list and an internal audit artifact containing audit logs. Furthermore, by replacing workflow definition files and structured rules, it supports multi-scenario compliance reviews and the generation of corresponding compliance review deliverables for mechanical drawings, electronic design packages, bills of materials, and standard operating procedures. Specifically, after unsuspending and completing the traversal of all nodes, the system generates compliance review deliverables based on the accumulated execution results in the shared state dictionary. The deliverables include a "problem rectification list" and a "compliance analysis report" for business personnel, as well as "internal audit artifacts" for the system backend. To meet the stringent traceability requirements of manufacturing quality systems (such as ISO / IATF), the system records the entire review process in a tamper-proof audit chain. It uses an append-only approach to record single-line audit logs (JSON Lines format). Each log entry contains review request metadata, a summary of the action performed, and a hash reference (prev_hash) pointing to the previous audit record. Based on this blockchain-like hash nesting structure, if any judgment result or file path in the historical record is tampered with, the subsequent hash chain will break completely. Through regular chain verification, the authenticity and integrity of the entire review record can be ensured.

[0029] Please see the appendix Figure 2 A human-machine collaborative manufacturing document compliance review system, comprising a processor and a memory, wherein the memory stores computer instructions, and when the computer instructions are executed by the processor, the following functional modules are implemented: The access request module is used to receive compliance review requests input by users and obtain workflow identifiers and manufacturing documents to be reviewed. The security filtering module is used to perform sensitive information detection on manufacturing documents and extract the document content of the manufacturing documents as initial processing data after the detection is passed. The workflow compilation module is used to read the corresponding workflow definition file based on the workflow identifier in the compliance review request, and parse and build it into an executable state diagram containing multiple processing nodes; The execution engine module is used to perform node traversal execution on the initial processing data based on the executable state diagram, call preset structured rules to automatically determine and extract evidence of compliance defects, generate automatic review conclusions, and suspend the current automated workflow when a manual confirmation node is triggered. The human-computer interaction module is used to output and display evidence of compliance defects and automatic review conclusions. After receiving manual confirmation instructions from users, it triggers the execution engine module to resume the process until the traversal is complete. The output and audit module is used to comprehensively generate the final compliance review deliverables and write the compliance review request, compliance review deliverables, and execution trajectory information of the review process into the tamper-proof audit chain. Specifically, to support the above method, this invention also provides a human-machine collaborative manufacturing document compliance review system. This system mainly includes the following modules: a request access module for receiving heterogeneous input data; a security filtering module for executing S0-level keyword and regular expression interception; a workflow compilation module responsible for parsing YAML files; a node registration module for managing atomic processing interfaces; an execution engine module for maintaining shared state and driving engine flow; and a human-machine interaction module and an output and auditing module for completing visual interaction and generating final deliverables. These modules work collaboratively, enabling this invention to quickly adapt to compliance review scenarios for various types of manufacturing documents, such as machinery, electronics, and procurement processes, without modifying the core system code, simply by adding workflow definition files and rule files.

[0030] Please see the appendix Figure 3 The workflow of this invention is defined using YAML metadata, dynamically compiled into an executable state graph by the compiler, and then driven by the executor. The compiler automatically discovers node implementations from the node registry and binds node-level configuration parameters to processing functions. The executor maintains a shared state dictionary, through which nodes pass intermediate results.

[0031] Please see the appendix Figure 4This invention incorporates security filtering layers at both the request entry point and the result output, employing a dual-engine detection mechanism of "keyword matching engine + regular expression engine." In the specific detection process, the system first receives the input text or request summary, then performs keyword scanning, followed by regular expression scanning, and finally summarizes the hit results to form a unified SOCheckResult. When any keyword or regular expression pattern is matched, the system determines that the current content triggers a sensitive information rule and returns a blocking conclusion, instead of continuing into the subsequent workflow compilation or large language model processing chain. The purpose of this is to move the security control point forward to the entry point, preventing sensitive content from spreading in subsequent nodes.

[0032] The advantages of this detection algorithm are twofold: firstly, keyword matching can quickly cover sensitive expressions at the business semantic level; secondly, regular expressions can reliably capture standard-formatted identity numbers and account information. The combination of these two features enables the system to identify both sensitive business information in the manufacturing industry, such as "quotations and profits," and highly structured sensitive personal information, thus adapting to the actual needs of "business sensitivity + personal sensitivity" coexisting in manufacturing document review scenarios.

[0033] Please see the appendix Figure 5 This invention employs a two-level routing collaboration mechanism. The first level involves the user explicitly selecting the workflow type in the interactive interface, such as mechanical_review, dcu_review, bom_review, or sop_review. The second level involves the Agent or system validator performing a validity check on the selection, that is, combining the format, quantity, directory structure of the uploaded files, and the requirements declared in input_schema to determine whether the current selection matches the input data.

[0034] Please see the appendix Figure 6 The present invention preferably adopts a review strategy of "rule priority and human-machine collaboration as a fallback". That is, in the automatic link, the rule_checker is executed first to match the structured rules, then the evidence_collector solidifies the evidence, and the decision_writer forms the conclusion. Only when more natural language expression is needed or the rules cannot cover the scenario will the template / llm / fallback routing or manual confirmation stage be entered. In this way, more than 80% of the standardized review can be completed automatically by the rules, while the human will only check the key decisions.

[0035] Please see the appendix Figure 7 The diagram contains the following modules / elements: YAML workflow, compiler, state diagram, executor, shared state, and deliverables.

[0036] Arrow / connection relationship: YAML → Compiler → State Diagram → Executor → Deliverable. Shared state is located in the center of the executor and can be read and written by each node.

[0037] Please see the appendix Figure 8 The diagram contains the following modules / elements: the horizontal axis represents four review scenarios (mechanical / electronic / BOM / SOP), the vertical axis represents the core capabilities of this invention (rule priority, security isolation, manual confirmation, audit chain, zero-code extension), and the arrows / connections: it is recommended to use tables or matrices instead of complex arrows; the coverage relationships are marked with checkmarks in the corresponding cross cells. This diagram is used to illustrate that this invention does not only cover a single file type, but also covers multiple manufacturing document review scenarios.

[0038] In summary, this invention provides a human-machine collaborative method and system for compliance review of manufacturing documents. By reading the corresponding workflow definition file based on the workflow identifier and dynamically constructing an executable state diagram, it achieves complete decoupling between business process definition and underlying execution code. This allows the system to quickly adapt to review scenarios of various heterogeneous manufacturing documents, such as mechanical drawings, electronic design packages, and BOM lists, without modifying the core code, simply by adding or modifying configuration files. The development cycle for a new type of manufacturing document review process is shortened from an average of 3-6 months under the traditional hard-coded model to just a few hours to a few days by writing workflow definition files and structured rule files. This reduces the system's development and maintenance costs and shortens the delivery cycle for new scenarios. Furthermore, by writing records including compliance review requests, review deliverables, and execution action trajectories into an anti-tampering audit chain using a single-line log with an appended hash reference value, a verification system based on chained hash comparison is established. This ensures that no operation throughout the entire review lifecycle can be silently overwritten or tampered with, fully meeting the stringent audit closed-loop requirements of the manufacturing quality management system for clear review responsibility and complete process traceability.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method of human-in-the-loop manufacturing industry document compliance review, the method comprising: Includes the following steps: S1. First, receive a compliance review request input by the user. The compliance review request includes at least a workflow identifier and a manufacturing document to be reviewed. S2. Subsequently, sensitive information detection is performed on the manufacturing documents. After determining that the manufacturing documents pass the sensitive information detection, the document content of the manufacturing documents is extracted as initial processing data. S3. Based on the workflow identifier in the compliance review request, read the corresponding workflow definition file, and parse the workflow definition file to construct an executable state diagram for performing compliance review. The executable state diagram contains multiple processing nodes. S4. Based on the executable state diagram, the initial processing data is traversed and executed node by node. During the traversal and execution of the processing nodes, preset structured rules are called to automatically determine the initial processing data in order to extract compliance defect evidence in the manufacturing documents and generate automatic review conclusions based on the compliance defect evidence. S5. During the process of traversing and executing the processing nodes, when a preset manual confirmation node in the workflow definition file is triggered, the current automated workflow action is suspended, and the compliance defect evidence and the automatic review conclusion are output and displayed to receive manual confirmation instructions from users based on human-machine collaboration feedback. The workflow definition file supports inserting one or more manual confirmation nodes between any processing nodes. The position and number of manual confirmation nodes can be configured by modifying the metadata of the workflow definition file without modifying the underlying execution code. S6. Upon receiving the manual confirmation instruction, resume the flow of the executable state diagram until the traversal is complete, combine the automatic review conclusion with the manual confirmation instruction to generate the final compliance review deliverable, and write the compliance review request, the compliance review deliverable, and the execution trajectory information of the review process into the anti-tampering audit chain.

2. The human-machine collaborative manufacturing document compliance review method of claim 1, wherein, The specific content of receiving user input for compliance review requests in step S1 includes: Receive the process type selected by the user in the interactive interface and use it as the workflow identifier; Based on the format and quantity of the uploaded manufacturing documents and the preset workflow input constraints, the workflow identifier is validated for reasonableness. Sensitive information detection of the manufacturing documents can only be triggered after the validation matches. The reasonableness validation includes: matching the file format of the uploaded documents with the accepted_formats declared in the input_schema in the workflow definition file; and validating the quantity of uploaded documents against the required field list declared in the input_schema.

3. The human-machine collaborative manufacturing document compliance review method of claim 1, wherein, The specific steps for performing sensitive information detection on the manufacturing documents in step S2 include: A business keyword matching engine and a regular expression engine are used to perform dual recognition and detection on the content of the manufacturing documents. If the blocking rule is met, the review process will be terminated; if the desensitization rule is met, the combination of personal sensitive information and business sensitive words will be replaced while preserving the context structure.

4. The human-machine collaborative manufacturing document compliance review method according to claim 1, characterized in that, The specific steps for extracting the document content of the manufacturing documents as initial processing data in step S2 include: The graphic and textual parameter information of the manufacturing documents is extracted using optical character recognition or document parsing methods to generate a structured summary, which is then used as the initial processing data. During subsequent traversal and execution of the processing nodes, the system only inputs the extracted initial processing data into the large language model node, prohibiting the original manufacturing documents from entering the large language model node, so as to achieve secure isolation of the source files.

5. The human-machine collaborative manufacturing document compliance review method according to claim 1, characterized in that, Step S3, which involves parsing the workflow definition file and constructing an executable state diagram for performing compliance reviews, specifically includes: The underlying atomic processing nodes are automatically scanned and discovered through a preset node registry. The configuration parameters in the workflow definition file, written in metadata format, are dynamically bound to the corresponding atomic processing nodes to generate the executable state diagram that transmits intermediate results using a shared state dictionary.

6. The human-machine collaborative manufacturing document compliance review method of claim 1, wherein, The automatic determination of the initial processed data by invoking preset structured rules in step S4 specifically includes: The structured rules are executed first to perform deterministic matching on the initial processed data, and to locate and extract the compliance defect evidence with a precise location and rule number; When the structured rules cannot cover the ambiguous semantic scenarios in the initial processed data, the large language model node is triggered by conditional routing to perform semantic supplementary interpretation, and the result is merged with the structured rule judgment result to generate the automatic review conclusion.

7. The human-machine collaborative manufacturing document compliance review method of claim 1, wherein, The specific steps in step S5, which involve receiving a user's manual confirmation instruction based on human-machine collaboration feedback, include: The front-end interactive interface highlights the evidence of compliance deficiencies and the corresponding risk locations, and provides a decision interface for users to perform approval or rejection operations; The user's actions on the decision interface are captured as the manual confirmation instruction and updated in the shared state dictionary of the executable state graph to release the suspended state.

8. The human-in-the-loop manufacturing documentation compliance review method of claim 1, wherein, Specifically, step S6, which involves writing the compliance review request, the compliance review deliverables, and the execution trajectory information of the review process into the tamper-proof audit chain, includes: Single-line audit logs are recorded using an append-only method. Each audit log entry contains metadata about the current compliance review request, the action performed, and a hash reference to the previous audit record. The integrity of the entire lifecycle operation record is verified by chaining the hash reference values.

9. A human-machine collaborative manufacturing document compliance review method according to claim 8, wherein, The steps for combining the automated review conclusions with the manual confirmation instructions to generate the final compliance review deliverable are further defined as follows: The compliance review deliverables include at least a user-facing issue rectification list and an internal audit artifact containing the audit logs; Furthermore, by replacing the workflow definition file and structured rules, it supports multi-scenario compliance review of mechanical drawings, electronic design packages, bills of materials, and standard operating procedures, as well as the generation of corresponding compliance review deliverables.

10. A human-machine collaborative manufacturing industry document compliance review system, characterized in that, According to any one of claims 1-9, a human-machine collaborative manufacturing document compliance review method, the system includes a processor and a memory, wherein the memory stores computer instructions, and when the computer instructions are executed by the processor, the following functional modules are implemented: The access request module is used to receive compliance review requests input by users and obtain workflow identifiers and manufacturing documents to be reviewed. A security filtering module is used to perform sensitive information detection on the manufacturing documents and extract the document content of the manufacturing documents as initial processing data after the detection is passed. The workflow compilation module is used to read the corresponding workflow definition file based on the workflow identifier in the compliance review request, and parse and construct it into an executable state diagram containing multiple processing nodes; The execution engine module is used to perform node traversal execution on the initial processing data based on the executable state diagram, call preset structured rules to automatically determine and extract evidence of compliance defects, and generate automatic review conclusions. And when a manual confirmation node is triggered, the current automated workflow action is suspended; The human-computer interaction module is used to output and display the compliance defect evidence and the automatic review conclusion. After receiving the manual confirmation instruction from the user, it triggers the execution engine module to resume the process until the traversal is completed. The output and audit module is used to comprehensively generate the final compliance review deliverables and write the compliance review request, the compliance review deliverables, and the execution trajectory information of the review process into the tamper-proof audit chain.

Images