An AI agent and skill package mechanism-based BOM automatic management method and system

Abstract

The application discloses a kind of BOM automation management method and system based on AI intelligent agent and skill package mechanism, it is related to manufacturing digital management technical field.The method includes: material list management field knowledge is encapsulated as the skill package using three levels progressive loading architecture, wherein metadata layer is always loaded, specification document layer is loaded on demand, resource and code layer is called on demand and does not enter context window;At least including BOM creation, checking, change management and cross-system synchronization four function modules Skill package set is built;Through skill registration framework, skill package is hung to AI intelligent agent;Through ReAct reasoning cycle, the automation execution of BOM management task is realized.The application also sets up environment isolation, permission minimization and operation auditable etc.Safety protection mechanism, solve the problem such as knowledge difficult to be structured in traditional BOM management Reprocessing, change response lag, cross-system coordination difficulty and material data non-standardization, realize the improvement of BOM management efficiency and data quality.

Description

Technical Field

[0001] This invention relates to the field of digital management technology in manufacturing, and more specifically, to a method and system for automated BOM management based on AI agents and skill packages. Background Technology

[0002] BOM (Bill of Materials) is the core foundational data for manufacturing informatization, permeating the entire manufacturing business process, including product design, process planning, production planning, material procurement, cost accounting, and after-sales service. In the process of digital transformation in traditional manufacturing, BOM management commonly faces the following technical challenges: The rationality and accuracy of a BOM structure heavily rely on the domain experience of senior process engineers. This knowledge is stored implicitly within individuals, lacking a systematic digital expression and transmission mechanism. When personnel leave or their positions change, the continuity and consistency of BOM maintenance are difficult to guarantee. Existing BOM management tools typically store BOM information as static records, failing to structure and automatically retrieve process knowledge, design specifications, and management experience.

[0003] In industries such as food, fast-moving consumer goods, and electronics manufacturing, Bill of Materials (BOM) changes are frequent, involving multiple dimensions of changes such as raw material replacement, formula adjustments, specification switching, and supplier changes. Traditionally, BOM changes rely on manual initiation, offline multi-departmental review, and manual updates to each system, with change cycles typically measured in days or weeks. Delayed change response directly leads to a chain of problems, including production plan disruptions, material stagnation or shortages, and delivery delays.

[0004] In the IT architecture of manufacturing enterprises, BOM data needs to flow across multiple heterogeneous systems such as PLM (Product Lifecycle Management), ERP (Enterprise Resource Planning), MES (Manufacturing Execution System), and WMS (Warehouse Management System). These systems are provided by different vendors, and their data models, field definitions, and interface protocols are incompatible. Traditional solutions use hard-coded point-to-point API connections, which suffers from long development cycles, high coupling, and high maintenance costs; any system upgrade may cause the interface to fail.

[0005] Different departments and suppliers use different coding rules, naming conventions, and specification descriptions for the same material. Problems such as duplicate material codes, misspelled names, and ambiguous specification parameters are common in BOMs, affecting the accuracy of Material Requirements Planning (MRP) calculations and the consistency of data across systems.

[0006] Based on a search of existing patent literature, the following related technical solutions have been disclosed in the field of BOM management: Chinese patent CN119887086A discloses a method for managing model parameters and bill of materials based on the Teamcenter platform. This method achieves BOM management through integrated data management, automated processes, access control, and version control. However, this solution relies on the Teamcenter platform, a specific PLM platform, and is incompatible with enterprise IT architectures outside of Teamcenter. Furthermore, it does not address the autonomous reasoning and decision-making capabilities of AI agents.

[0007] Chinese patent CN104778540A discloses a BOM management method for building materials equipment manufacturing, including steps such as establishing a unified material library, defining BOM set relationships, inputting BOM data, generating details, generating material plans, and managing changes. This solution systematizes the BOM management process, but it is still driven by manual operation and does not achieve intelligent automatic identification and processing.

[0008] Chinese patent CN120541086A discloses a method for managing and updating BOM and color databases, which updates the database after parsing the BOM file, TC table file, and color list into a readable state. This solution focuses on parsing and storing data in a specific format and does not involve using AI technology for intelligent decision-making or automated change control.

[0009] Chinese patent CN119066034B discloses a method, system, and BOM order update method for capturing BOM effective dates, which predicts BOM effective dates using an elliptical geometric model. This solution is effective in predicting effective dates, but its applicability is limited to a specific calculation model and does not cover the entire lifecycle management of BOMs.

[0010] In summary, the existing technologies have the following shortcomings: (1) BOM management knowledge has not been encapsulated in a structured manner and can be expressed in a reusable way; (2) Change processing relies on manual driving and has a large response delay; (3) Cross-system data synchronization adopts a hard-coded method, which is not flexible and scalable enough; (4) Material data standardization lacks intelligent automatic processing methods; (5) There is a lack of a complete technical solution for BOM full life cycle management that combines the autonomous reasoning ability of AI agents. Therefore, to address the above issues, a method and system for automated BOM management based on AI intelligent agents and skill package mechanisms is proposed. Summary of the Invention

[0011] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a BOM automated management method and system based on an AI agent and skill package mechanism, in order to solve the technical problems in the prior art such as the difficulty in structured reuse of BOM management knowledge, delayed change response, difficulty in data collaboration across heterogeneous systems, and low standardization of material data.

[0012] To achieve the above objectives, this invention provides an automated BOM management method based on an AI agent and skill package mechanism. This method encapsulates professional knowledge in the field of BOM management into skill packages that can be invoked on demand by the AI ​​agent. Automated management of the entire BOM lifecycle is achieved through a progressive loading mechanism and a ReAct inference loop. The method includes the following steps: Step S1: Encapsulate the professional knowledge in the domain of bill of materials management into multiple skill packages for AI agents to call. Each skill package adopts a three-level progressive loading architecture, which includes: a metadata layer that is always loaded when the agent starts, a documentation layer that is loaded on demand when a task is triggered, and a resource and code layer that is called on demand during the execution phase without entering the agent's context window. Step S2: Construct a set of functional modules that include at least the following: BOM creation skill module, BOM verification skill module, BOM change management skill module, and BOM cross-system synchronization skill module; Step S3: The skill package is attached to the AI ​​agent through the skill registration framework, so that the agent can discover and call the registered skills during runtime; Step S4: Drive the AI ​​agent to perform BOM management tasks through the ReAct inference loop. The ReAct inference loop includes: scanning the metadata layer of registered skills for intent recognition and skill matching, loading the description document layer of the matched skills to obtain the operation process and rules, calling resources and code layers to perform operations, and generating BOM data files and pushing them to the target business system for result output and synchronization.

[0013] Furthermore, in the three-tier progressive loading architecture in step S1: the metadata layer includes skill names and function descriptions, and the metadata layer token consumption for a single skill pack is less than 100 tokens; The documentation layer contains operation guidelines, workflows, and constraint rules. The documentation layer token consumption for a single skill pack is less than 5000 tokens, and it only occupies the agent context window during task execution. The resource and code layer contains executable scripts, reference documents, BOM structure templates, and validation rule sets. This layer of resources does not enter the agent context window; it is only loaded and run by the execution environment when used.

[0014] Furthermore, the BOM creation skill module in step S2 is used to automatically generate a structure tree of the manufacturing BOM based on the product design data, and its execution flow includes: Extract structured data from the design BOM from the CAD / PLM system; Convert the design BOM into a manufacturing view and add process virtual parts and auxiliary material information; Batch generate or verify material codes by referencing material coding rule files; Output manufacturing BOM data file.

[0015] Furthermore, the BOM verification skill module in step S2 is used to automatically verify the integrity, consistency, and compliance of BOM data, and its verification mechanism includes: Based on a preset rule set, the BOM data is automatically audited in batches to ensure the completeness of required fields, the uniqueness of material codes, and the correctness of parent and child item usage. Using knowledge graphs in the manufacturing sector, automatic error correction of material names and specifications is achieved through entity recognition and semantic matching; It also generates a structured verification report that identifies anomalies, anomaly types, and suggested corrections.

[0016] Furthermore, the BOM change management skill module in step S2 is used to process engineering change requests, and its execution flow includes: Receive change instructions and extract the change type, target material, and change parameters through natural language processing; Automatically retrieve all BOM versions containing the target material and assess the impact of changes on inventory, in-transit purchases, work-in-process orders, and costs; Perform batch modification of BOM master data and generate new version numbers while retaining the historical version traceability chain, and synchronize the update to related systems; Change notifications are automatically distributed according to preset rules to form a closed loop.

[0017] Furthermore, the BOM cross-system synchronization skill module in step S2 is used to realize the automatic synchronization of BOM data between heterogeneous systems such as PLM, ERP, and MES. Its synchronization methods include: For legacy systems that do not support standard APIs, intelligent screen understanding technology is used to visually recognize system interface elements and simulate manual operations to complete data reading and writing. The semantic understanding capabilities of large language models are used to identify the differentiated codes and field representations of the same material across different systems and to establish a coding mapping relationship table. It supports local deployment on enterprise servers or private cloud environments, ensuring that data does not leave the enterprise network boundaries.

[0018] Furthermore, the ReAct inference loop in step S4 specifically includes the following sub-steps: S4-1 Intent Recognition and Skill Matching: Matches the semantics of the user's task with the metadata layer information of all registered skills to filter out the most relevant skill package combinations; S4-2 Knowledge Loading: Read the contents of the description document layer of the matching skill and load it into the agent's context window to obtain the operation process, constraint rules and processing strategies; S4-3 Action Execution: Following the instructions in the documentation, invoke the resources of the matching skills and the executable scripts and resources in the code layer to perform BOM management operations; S4-4 Result Output and Synchronization: Generate a BOM data file, and after manual confirmation, call the BOM cross-system synchronization skill module to push it to the target business system.

[0019] Furthermore, the method also includes security protection steps: Implement network or physical layer isolation between the intelligent agent's operating environment and the production system. Deploy the intelligent agent on a dedicated server or independent virtual machine instance, and only open the network ports and access paths necessary for the task. Configure permissions to minimize the file system directories, application programming interfaces and network addresses accessible to the agent, granting only the minimum set of permissions necessary to complete the scheduled BOM management tasks; All operations performed by the intelligent agent during its operation are logged, and a second manual confirmation step is set up for irreversible operations to ensure that the final execution authority is in the hands of the operator.

[0020] Furthermore, the method also includes at least one of the following response mechanisms: BOM high-frequency change response mechanism: Automatically generate a new version number and retain a complete historical version traceability chain through version snapshot management; use AI to learn patterns from historical change data to predict BOM adjustment needs and generate pre-adjustment suggestions. Non-standardized material data handling mechanism: A data cleaning pipeline is preset in the BOM verification skill module, which uses rule matching and vector similarity calculation to automatically map non-standard material names and specifications to the enterprise's standard material master data entries. Cross-departmental collaborative response mechanism: The automatic notification function of the BOM change management skills module automates change approval and cross-system synchronization. When the cost change caused by the BOM change exceeds the preset threshold, an early warning notification is automatically generated and a manual intervention process is triggered.

[0021] This invention also provides an automated BOM management system based on an AI agent and skill package mechanism, comprising: The skill package building module is used to encapsulate knowledge in the bill of materials management domain into multiple skill packages using a three-tier progressive loading architecture. The three tiers include a metadata layer, a documentation layer, and a resource and code layer. Functional module unit, used to manage a set of functional modules including at least a BOM creation skill module, a BOM verification skill module, a BOM change management skill module, and a BOM cross-system synchronization skill module; The skill registration and mounting module is used to mount the skill package to the AI ​​agent instance through the skill registration framework; An automated execution module is used to drive an agent to perform BOM management tasks through a ReAct inference loop, wherein the ReAct inference loop includes intent recognition and skill matching, knowledge loading, action execution, and result output and synchronization. The security protection module is used to implement runtime environment isolation, permission minimization control, and operation auditing.

[0022] Compared with the prior art, the present invention has the following beneficial effects: Knowledge assetization: Encapsulate discrete BOM management domain knowledge, process experience, and operating procedures into structured, AI-reusable skill packages, enabling the explicit expression of tacit knowledge and cross-project reuse, reducing reliance on individual senior personnel.

[0023] Agile Change Management: Through intelligent change management skills, change requests are automatically parsed, the scope of impact is automatically assessed, multiple systems are linked for updates, and notifications are closed-loop, reducing BOM change processing time from the traditional manual days to automated minutes.

[0024] Cross-system data consistency: By using cross-system synchronization skills, combined with semantic mapping and visual understanding technologies, automatic alignment and consistency verification of BOM data between heterogeneous systems such as PLM, ERP, and MES are achieved, eliminating data silos.

[0025] Material data quality improvement: Through an automatic verification mechanism that combines rule engine and knowledge graph, the integrity and standardization issues of material data are systematically detected and corrected, thereby improving the quality of BOM data.

[0026] Operational security assurance: Through a three-layer security mechanism of environment isolation, privilege minimization, and auditable operations, data security and operational compliance are ensured during the BOM automated management process.

[0027] Token efficiency optimization: Through a three-level progressive loading architecture, the agent does not occupy additional context resources during non-BOM tasks, and only loads relevant knowledge on demand when the task is triggered, which significantly reduces the cost of calling large language models. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the system module composition of the present invention; Figure 2 This is a schematic diagram of the system workflow of the present invention; Figure 3 This is a flowchart illustrating the internal execution process of the BOM creation skill module of this invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] As attached Figures 1 to 3 The invention describes an automated BOM management method and system based on AI agents and skill pack mechanisms. In this description, it should be noted that the terms "BOM" refer to Bill of Materials, "EBOM" to Design Bill of Materials, "MBOM" to Manufacturing Bill of Materials, "PLM" to Product Lifecycle Management System, "ERP" to Enterprise Resource Planning System, "MES" to Manufacturing Execution System, "WMS" to Warehouse Management System, "CAD" to Computer-Aided Design System, "ECR / ECO" to Engineering Change Request / Engineering Change Order, "ReAct" to a cyclical reasoning pattern of reasoning and action, "Token" to the smallest semantic unit in a large language model processing text, and "skill pack" to a collection of functional modules encapsulating specific domain knowledge, operational procedures, and executable resources.

[0031] I. Implementation Architecture of the Overall Technical Solution The BOM automated management method proposed in this invention, based on AI intelligent agents and skill package mechanisms, extracts professional knowledge in the field of BOM management from the implicit experience of senior process engineers, encapsulates it into a structured skill package that can be called on demand by AI intelligent agents, and achieves automated management of the entire life cycle of BOM through a progressive loading mechanism and ReAct inference loop.

[0032] The implementation of the entire system includes a system construction phase and a task execution phase.

[0033] During the system construction phase, the corresponding steps are S1 to S3: First, the professional knowledge in the field of BOM management is packaged into multiple skill packages according to a unified three-tier architecture; Based on the business needs of the entire BOM management lifecycle, we have built a skill set that covers four major functional modules: BOM creation, BOM verification, BOM change management, and BOM cross-system synchronization. Then, these skill packages are attached to AI agent instances through a skill registration framework, enabling the agents to discover and invoke them at runtime.

[0034] During the task execution phase, corresponding to step S4: When a user initiates a BOM management task, the AI ​​agent autonomously completes the entire process from intent recognition and skill matching, knowledge loading, action execution to result output and synchronization through the ReAct inference loop. The entire operation is protected by a multi-layered security protection mechanism, corresponding to step S5. The security protection module runs through the entire operation process, controlling the agent's operating environment, operating permissions, and operation records. In addition, supplementary response mechanisms are provided for special scenarios in BOM management, corresponding to step S6.

[0035] II. The three-tiered progressive loading architecture of skill packs Each BOM management skill package is encapsulated using a unified three-tier progressive loading architecture. This architecture divides the knowledge content of the skill package into three tiers based on usage frequency and context window usage requirements. The three tiers work together through explicit invocation conditions and data isolation mechanisms.

[0036] First level – Metadata layer This layer is fully loaded into memory during agent initialization and remains resident in the agent's available skill index table. The metadata layer is a lightweight descriptive data structure, with each skill pack occupying approximately 100 tokens.

[0037] Taking the BOM creation skill package as an example, its metadata layer contains the following information: the skill name is "bom-creator", the function description is "automatically generate the structure tree of the manufacturing BOM based on product design data or ERP work orders", the applicable scenarios include "new product import, customer order BOM generation, BOM copying and derivation", and the triggering keywords include "BOM creation, generating MBOM, EBOM conversion, manufacturing BOM".

[0038] During the agent's startup phase, the system traverses all subdirectories under the skill pack's root directory, reads the metadata files in each subdirectory, and constructs a skill index table from the metadata of all registered skills. This index table remains resident in memory throughout the agent's entire runtime. This design makes the agent's initial state extremely lightweight; even if dozens of skill packs are registered in the system, the initial token consumption is only in the range of one or two thousand tokens.

[0039] Second level – Documentation layer This layer is loaded onto the context window only when the agent determines, through matching the first-level metadata, that a specific skill needs to be invoked. The documentation layer is a complete operation guide file (e.g., named SKILL.md), and the documentation for a single skill pack is limited to 5000 tokens.

[0040] Taking the BOM verification skills package as an example, its documentation contains the following structured content: The operation guide section details the complete sequence of steps for BOM verification, including the first step of performing rule engine verification, the second step of invoking semantic error correction for ambiguous items that the rule engine cannot determine, and the third step of summarizing and generating a verification report.

[0041] The standard workflow section clarifies the triggering conditions, prerequisites, input data format, and output data format for validation.

[0042] The constraint rules section defines the judgment criteria and boundary conditions for verification, such as the allowable deviation range of parent and child item usage, the legal character set of material codes, and a complete list of required fields.

[0043] The exception handling strategy section describes the alternative solutions to be taken when the verification script fails to execute, the knowledge graph query times out, or the BOM data format does not meet expectations.

[0044] The documentation layer is loaded into the agent's context window only during task execution. Once the task is complete, the layer's contents are released from the context window and do not occupy context resources for subsequent tasks.

[0045] Third level – Resources and code layer During task execution, this layer is accessed by the agent as needed, following the guidance of the documentation layer, through script calls or file reading. The resource and code layer contains the following types of executable resources and data files: Executable scripts, such as Python scripts for extracting design BOM data from CAD / PLM systems, scripts for performing EBOM to MBOM conversion, scripts for calling enterprise material master data APIs, and scripts for controlling browser automation frameworks, etc.

[0046] Reference documents, such as enterprise material coding standards and industry BOM management standard documents.

[0047] A BOM structure template, such as an Excel or JSON format MBOM template file, defines the standard field structure of the MBOM.

[0048] Data validation rule sets, such as rule configuration files in YAML format, explicitly define rule items such as mandatory field checks, code uniqueness checks, and usage calculation verification.

[0049] The core characteristic of this resource level is that it does not enter the agent's context window, and the token consumption is almost zero. When the agent needs to call a script during the action execution phase, it generates code instructions to execute the script. The script's execution process and results are completed in an external execution environment. The agent determines whether the operation was successful by reading the standard output information and status code returned after the script execution.

[0050] Collaborative workflow between the three levels: The specific triggering conditions and switching logic between the three levels are as follows: When the system starts, only the first level of all skills is loaded into memory, a skill index table is built, and the agent is in a lightweight standby state.

[0051] When a user task arrives, the agent matches the task semantics with the metadata layer information of each skill in the skill index table. This process consumes very few tokens. Once a target skill is matched, the second-level loading condition is triggered, and the agent loads the skill's documentation into the context window via a file read command.

[0052] Once the documentation is loaded, the agent obtains a complete operational blueprint. When executing specific operations, the agent, according to the steps defined in the documentation, calls the specific scripts and data files at the third level as needed by generating code instructions. These resources run in an external execution environment without entering the context window.

[0053] Once the task is completed, the second-level content can be released from the context window, while the first-level index remains resident.

[0054] Throughout the process, the switching between the three levels is driven by the state judgment in the ReAct inference loop: after each action, the agent observes the result and decides whether to continue to call the resources of the third level, fall back to the second level to check the exception handling strategy, or end the task and release the resources.

[0055] III. Registration and Mounting of Skill Packs During the system deployment phase, the skill packages constructed above are mounted to AI agent instances through the skill registration framework.

[0056] In its implementation, the skill registration framework provides a file system registrar. This registrar reads a configuration file, which specifies the root directory path of the skill package. The registrar traverses all subdirectories under this path, searches for metadata files in each subdirectory, extracts the skill name, function description, and trigger keywords, and builds a skill index table in memory.

[0057] After registration, metadata summary information of currently available skills is injected into the system prompts of the AI ​​agent, enabling the agent to perceive the capability range of all available skills during subsequent reasoning.

[0058] When an agent receives a user task, it first performs a matching search in the skill index table within its context, matches the corresponding skill according to the task semantics, then loads the description document of the matched skill through a file reading tool, and then calls the third-level resources according to the guidance of the description document.

[0059] This registration and invocation mechanism allows skill packs to be dynamically added, updated, or removed without modifying the agent's core code, achieving a high degree of scalability.

[0060] IV. Implementation of the BOM Creation Skills Module The BOM creation skill module is used to automatically generate the structure tree of the manufacturing BOM based on upstream design data. The following explanation uses an electronics manufacturing company as an example.

[0061] The company's R&D department uses mainstream CAD software to complete the printed circuit board design and then uses the software's export function to generate a structured data file conforming to industry standards. This file contains complete circuit board design information, including component list, reference designators, package type, quantity, and hierarchy. The user provides the storage path of this design file as a task parameter to the agent, which then reads the design file from the specified path or obtains it by calling the PLM system interface.

[0062] In the intent recognition and skill matching stage, the agent compares the semantics of the user's task with the metadata of all skills in the skill index table, and identifies the skill that best matches the task as bom-creator through semantic similarity calculation.

[0063] During the knowledge loading phase, the agent calls the file reading tool to load the documentation in the bom-creator skill package directory into the context window and obtain the complete operation process of the skill: the first step is to parse the design file to extract the EBOM structure, the second step is to convert the EBOM to MBOM according to the conversion rules, the third step is to generate and verify the material code, and the fourth step is to output the MBOM data file.

[0064] During the action execution phase, the agent generates executable code and calls third-level script resources step by step, following the instructions in the documentation.

[0065] The first step is for the agent to invoke the design data parsing script. This script receives the design file path as input, uses XML path language to locate and extract all component nodes from the input file, traverses each node to obtain the component number, functional description, package type, usage, reference designator, and parent-child hierarchy, and constructs the EBOM (Extended Embedded Component Object) structure tree in memory. After the script executes, it outputs the EBOM structure tree to a temporary file in JSON format and returns the execution status code and statistics (including total number of components, hierarchy depth, etc.) via standard output.

[0066] The second step involves the agent invoking the EBOM to MBOM conversion script. This script reads the EBOMJSON file generated in the first step, and simultaneously reads the preset conversion rule file from the third level. The conversion rule file defines multiple conversion logics in a structured manner. Example rules include: if the component type is a surface mount resistor and there are no child nodes in the level, then a virtual parent component is automatically created; the fastener pairing rule is that each screw is automatically matched with the corresponding number of nuts and washers. The script executes these rules item by item, adding process virtual component nodes, auxiliary material nodes, and packaging material nodes to the BOM structure tree in memory, completing the EBOM to MBOM conversion.

[0067] The third step involves the agent invoking the material code generation script. This script reads the enterprise material coding rule file in the third level, which defines the code segment structure. The script iterates through each new material node in the MBOM structure tree, matching the code segment definition to its material category to generate a unique material code. For existing codes inherited from the EBOM, the script performs uniqueness verification by calling the enterprise material master data system API. If the API returns that the code already exists, a replacement code is automatically generated according to the conflict handling rules, and the code mapping relationship is recorded.

[0068] Fourth, after the above operations are completed, the script serializes the MBOM structure tree in memory into a JSON data file and generates a report listing the total number of materials processed, the number of newly generated codes, and the statistical information such as detected anomalies.

[0069] During the result output and synchronization phase, the agent presents the generated MBOM JSON file and report to the user for confirmation. After the user confirms that there are no errors, if it is necessary to push the MBOM to the downstream system, the agent further calls the BOM cross-system synchronization skill module to complete the push.

[0070] V. Implementation of the BOM Verification Skills Module The BOM verification skills module is used to automatically verify the integrity, consistency, and compliance of BOM data. The following detailed description, in conjunction with examples, illustrates this process.

[0071] Once the MBOM is generated, the agent automatically invokes the bom-validator skill according to the preset subsequent steps in the task chain. During the knowledge loading phase, the agent loads the skill's documentation into the context window, learning that the validation pipeline consists of three steps: rule engine validation, knowledge graph error correction, and report summary generation.

[0072] During the action execution phase, the agent performs the following operations in sequence.

[0073] Step 1 -- Rule Engine Verification.

[0074] The agent invokes the rule engine script in the third layer, which loads a set of validation rules in YAML format. The validation rule set defines multiple validation rules in a structured manner, mainly including: Completeness check of required fields: Traverse each material node in the BOM and check whether there are any empty values ​​in key fields such as material code, material name, specifications, quantity, and unit.

[0075] Material code uniqueness verification: Extract all material codes from the BOM and check for duplicate codes using a hash table data structure. If the same code appears in two or more different material rows, it is recorded as an anomaly.

[0076] Verification of the correctness of the parent-child item usage calculation: For each assembly node, summarize the unit usage of all its direct child nodes and verify the equality with the assembly quantity of that assembly node in the parent.

[0077] Encoding format compliance check: Based on the legal character set and length range defined in the encoding rule file, perform regular expression matching verification on each material code.

[0078] The rule engine script executes the above rules one by one, and records data items that do not conform to the rules into the exception list. Each record includes the location of the exception, the exception type, the original data, and the exception level.

[0079] Step 2 - Knowledge Graph Error Correction.

[0080] For semantic-level issues that the rule engine cannot determine, the agent calls a semantic error correction script for processing.

[0081] The script first reads the enterprise material knowledge graph interface configuration in the third level. This knowledge graph is built during the system's offline phase. It constructs a graph structure containing material entities, attribute values, and interrelationships by collecting historical BOM data, material master data, and industry standard material classification systems.

[0082] Secondly, the script concatenates the material name and specifications to be verified into a text string, and uses text embedding technology to convert it into a vector representation.

[0083] Then, a nearest neighbor search is performed in the standard material vector index of the knowledge graph, and the cosine similarity between the query vector and each standard material vector is calculated. If the standard material entry with the highest similarity is found and exceeds the preset threshold, the script automatically replaces the original non-standard description with the standard description and records it as automatic correction in the anomaly list; if the highest similarity is still below the threshold, the material is marked as awaiting manual confirmation, and the top three candidate standard entries with the highest similarity and their similarity values ​​are attached.

[0084] Step 3 - Report compilation and generation.

[0085] The script aggregates all exception records generated in the first two steps and generates a structured validation report according to a unified JSON structure. Each exception record in the report includes the following fields: exception identifier, exception type, original data content, suggested correction method, and exception level. This validation report is then presented to the user for review.

[0086] VI. Implementation of the BOM Change Management Skills Module The BOM change management skills module is used to process engineering change requests, automatically assess the scope of the change's impact, and update the relevant BOM accordingly.

[0087] The operation process of this module is divided into four steps: change request parsing, impact scope analysis, change execution, and notification closure. The following detailed explanation is provided with reference to an example.

[0088] Suppose a customer requests a change to replace material A in a product with material B. The user issues an instruction to the agent to process this change request. The agent matches the `bom-change-manager` skill through intent recognition, loads its documentation, and obtains the four-stage change management workflow.

[0089] Step 1 -- Change Request Parsing.

[0090] The agent utilizes the natural language understanding capabilities of a large language model to extract structured change elements from the change instruction text. The extracted results contain the following key-value pairs: change type is material replacement, source material code, target material code, and applicable scope is the corresponding product model. This structured data is saved as a JSON object for use in subsequent steps.

[0091] Step 2 – Scope of impact analysis.

[0092] The agent invokes the impact scope analysis script in the third level. This script connects to the enterprise's BOM version database and performs the following queries and analysis operations: BOM Impact Scope Query: The script executes a recursive query to retrieve all BOM node records containing the target material from the BOM structure table, and traces upwards along the parent-child relationship chain to the top-level product model and BOM version number to form a list of affected BOM versions.

[0093] Inventory and Procurement Impact Query: The script uses a pre-defined read-only interface to the ERP system database to query the current inventory of the target material in the inventory table, the quantity of in-transit purchase orders and their expected delivery dates in the purchase order table, and the information on incomplete production orders in the production work order table.

[0094] Cost impact calculation: The script obtains the unit standard cost of the source material and the target material through a preset cost accounting interface, calculates the unit cost difference of the replacement operation, multiplies it by the work-in-process quantity of the affected products, and obtains the total cost impact amount and change ratio.

[0095] The above results are compiled to generate a structured impact assessment report, which is then presented to the user for confirmation.

[0096] Step 3 - Implement the changes.

[0097] After user confirmation, the agent invokes the change execution script. This script performs the following operations within the same database transaction: marks the activation status of the currently active version as historical and records the archive timestamp; creates a new version record using the current version data as a template, generates an incrementing version number, sets the parent version ID, and establishes a traceability chain; performs material replacement in the new version's BOM structure and updates relevant attributes; and inserts a traceability record into the change traceability table. All these operations are completed within the same database transaction, ensuring data consistency.

[0098] Step 4 – Notification Closure.

[0099] After the change is successfully executed, the agent invokes the message push script. This script reads the preset notification rule configuration file, automatically matches the list of departments that should be notified based on the change type, generates a formatted change notification form, and automatically pushes the change notification to the corresponding department groups through the enterprise messaging middleware, completing the change loop.

[0100] VII. Implementation of the BOM Cross-System Synchronization Skills Module The BOM cross-system synchronization skill module is used to achieve automatic synchronization and consistency verification of BOM data between heterogeneous systems such as PLM, ERP, and MES. The following detailed explanation uses an example from an industrial equipment manufacturing enterprise.

[0101] The company's IT architecture includes multiple systems: the PLM and ERP systems provide standard REST API interfaces, while the older MES system only provides a graphical user interface and lacks a programmable API interface. Therefore, cross-system synchronization requires different interaction strategies for different types of systems.

[0102] For systems with standard APIs: For PLM and ERP systems, the agent invokes the API interaction script in the third layer. This script pre-configures the API endpoints for each system, including the base API URL, authentication method, and request format.

[0103] When retrieving EBOM data from the PLM system, the script constructs an HTTP GET request, carries an authentication token to access the PLM's BOM export API endpoint, parses the returned JSON-formatted EBOM data, and saves it to a local temporary file. When pushing MBOM data to the ERP system, the script constructs an HTTP POST request, uses the MBOM JSON data as the request body, carries an authentication token to access the ERP's BOM import API endpoint, parses the returned response status code and synchronization result identifier, and confirms whether the synchronization was successful. If the request fails, the script retryes according to a preset retry policy.

[0104] For legacy systems that do not support standard APIs: For older MES systems that only provide a graphical user interface, the intelligent agent uses intelligent screen understanding technology for data interaction. The implementation of this technology includes three stages: model preparation, operation sequence generation, and real-time execution.

[0105] First, model preparation.

[0106] During the preparation phase before system deployment, technicians captured screenshots of the MES system's BOM import interface, covering various operational states and interface changes. Annotation tools were used to manually label the target interface elements in the screenshots, including element category and location. The pre-trained object detection model was then fine-tuned using the labeled dataset. After training, the model accurately identified various operational elements and their precise coordinates in the MES system's import interface. The model file and category label mapping file are stored in the third-level resources of the skill package.

[0107] Second, operation sequence generation.

[0108] During the real-time synchronization phase, the agent first reads the skill's documentation to obtain the preset sequence of operation steps required to complete BOM import in the MES system. An example sequence is: click the BOM Management menu, click the BOM Import submenu, enter the path to the file to be imported in the file path input box, select Manufacturing BOM in the Material Type drop-down menu, click the Submit button, and observe the returned results. The documentation pre-defines the target element descriptions for each operation step, and the agent uses these steps as an operation blueprint.

[0109] Third, it is executed in real time.

[0110] The agent invokes the screen comprehension synchronization script in the third layer. This script controls the browser automation framework to launch a controlled browser instance, navigate to the MES system's BOM import page, and execute the operations step by step according to the sequence.

[0111] In each step, the script first takes a screenshot of the visible area of ​​the current browser window, sends the screenshot image data into the trained interface element detection model for inference, and the model outputs a list of all detected interface elements, each containing a category label and bounding box coordinates.

[0112] Then, the script searches for matching elements in the detection results based on the target element description of the current step. After finding the target element, the script calculates its center point position based on the bounding box coordinates and simulates a mouse click operation at that coordinate position using the browser automation framework.

[0113] Based on the response type of the interface after the click operation, the script executes the corresponding subsequent sub-operations: if a system file selection dialog box pops up, the local path of the MBOM file to be uploaded is filled in through the file selection interface of the automation framework and confirmed; if an option needs to be selected from a drop-down menu, the menu option element is detected again after the drop-down menu is expanded, the target option is located and clicked; if text needs to be entered in an input box, the input box element is located and keyboard input is simulated.

[0114] After each step is completed, the script briefly waits and observes the page changes to determine if the current step was executed successfully. If the expected result is detected, it continues to the next step; if the expected result is not detected or an error message appears, it retryes or logs the error according to the error handling strategy in the documentation. The entire process completely simulates the operation sequence of a human operator performing BOM import on the MES system interface, without requiring any code-level modifications to the MES system.

[0115] Cross-system material coding semantic mapping: During synchronization, different systems may encode the same material differently. To address this issue, the third level of the skill pack includes semantic mapping scripts and mapping relationship tables.

[0116] The semantic mapping script periodically exports the latest material master data tables from each business system and loads them into memory. The script utilizes the semantic matching interface of a large language model to compare the semantic similarity of material names and specifications across systems. For matching pairs with a confidence level higher than a preset threshold, the script records them in a cross-system material code mapping table.

[0117] During synchronous execution, the script first queries the mapping table for encoding translation; if no existing mapping exists, it calls the semantic matching interface in real time to try to establish a new mapping relationship; if a match still cannot be found, it is marked as requiring manual confirmation of the mapping.

[0118] Localized security deployment: The entire BOM cross-system synchronization skill module and all its required dependencies are deployed in containers on enterprise servers or in private cloud environments. All data interactions are conducted through the enterprise intranet. Firewall rules ensure that the agent server cannot access the external internet, ensuring from a network architecture perspective that BOM data never leaves the enterprise network boundary during processing and transmission.

[0119] VIII. The Complete Execution Process of the ReAct Inference Loop The ReAct inference loop comprises four sub-steps: intent recognition and skill matching, knowledge loading, action execution, and result output and synchronization. The following example, a complete BOM creation task, details how each sub-step works collaboratively. It's important to note that the ReAct inference loop is not a simple linear execution, but rather an inner iterative loop of inference-action-observation-adjustment. The agent determines the execution result of each step by reading the standard output information and status codes returned after script execution.

[0120] S4-1 Intent Recognition and Skill Matching After the user sends a task instruction to the agent, the agent first converts the user instruction text into an embedding vector. Then, the agent iterates through all entries in the skill index table in memory, and for each registered skill, it concatenates its skill name, function description, and trigger keywords into a skill description text, which is also converted into an embedding vector.

[0121] Next, the cosine similarity between the user instruction vector and each skill description vector is calculated. For the BOM creation task, the user instruction has the highest similarity score with the bom-creator skill, so the agent determines that the current task requires invoking the bom-creator skill. Simultaneously, the agent reads the associated skill field from the bom-creator's metadata. This field indicates that BOM validation is typically required after BOM creation, so the agent marks the bom-validator as a subsequent skill in the task chain.

[0122] S4-2 Knowledge Loading After the agent determines to invoke the bom-creator skill, it generates a file reading command to read the documentation in the skill package directory. The documentation content is fully loaded into the agent's current context window. The agent parses the content structure of the documentation to obtain the task's operation stages, the input and output requirements for each stage, the script file paths and parameter descriptions, as well as possible exceptions and corresponding handling strategies for each stage.

[0123] Operation S4-3 Following the instructions in the documentation, the agent begins to generate executable code and invoke third-level resources step by step. After each action is executed, the agent reads the standard output information and status code returned by the script, determines whether it was successful, and decides on the next action. This process of action-observation-determination-next action constitutes the inner iteration of the ReAct loop.

[0124] The specific execution process is as follows: The agent generates code that calls the design data parsing script, passing the design file path as a parameter. After execution, the script returns a status code of 0 (indicating success) and statistical information. Upon receiving status code 0, the agent determines that parsing is complete and proceeds to the next step.

[0125] The agent generates code to call the EBOM to MBOM conversion script. After the script executes, it returns a status code of 0. The agent reads the status code 0, determines that the conversion was successful, and continues to the next step.

[0126] The agent generates code that calls the material code generation script. If the script returns a connection timeout error when calling the material master data API, the agent detects this exception by reading the standard output information. Based on the pre-defined exception handling strategy in the documentation (retrying after a preset time interval, up to a preset number of retries if the API call fails), the agent waits and then re-executes the code generation operation. On the second execution, the API responds normally, the script completes the code generation, and returns a status code of 0. The agent reads the success status and determines that this step is complete.

[0127] The agent generates code that serializes the MBOM structure tree and writes it to a JSON file. After successful file writing, the agent confirms the output file has been generated by reading standard output information.

[0128] S4-4 Result Output and Synchronization Once all operations are successfully completed, the agent summarizes the execution process and results information and presents it to the user in a readable report format. The user views the MBOM data preview and report content, and clicks "Confirm Publish" after confirming that everything is correct.

[0129] After user confirmation, the agent automatically invokes the BOM-Sync skill module based on the subsequent skills marked in the task chain. The agent takes the MBOM file path and a list of target systems as input, loads the BOM-Sync skill documentation, and sequentially pushes the MBOM data to each target business system. After the push is complete, the agent records the task completion status and sends task completion information to the user.

[0130] IX. Implementation of Safety Protection Mechanism During deployment and operation, this system implements a three-tiered security protection mechanism: operational environment isolation, permission minimization, and operational auditability. These mechanisms work together to form a defense-in-depth system.

[0131] runtime environment isolation mechanism: At the physical network topology level, the AI ​​agent server is deployed in an independent virtual local area network, and is logically isolated from the enterprise's daily office network and core production system network through firewall devices.

[0132] The firewall outbound rules are set as follows: Only allow the agent server to initiate outbound connections to specific ports of the BOM database server; only allow outbound connections to specific ports of the API gateways of the PLM and ERP systems; only allow outbound connections to specific ports of the message middleware. The destination IP address and port of all outbound connections must be explicitly listed in the whitelist.

[0133] The firewall inbound rules are set to: block inbound connection requests to the agent server from any external network, open only the management ports required by the management console, and restrict connections to only those from the specified management jump server IP address.

[0134] The intelligent agent server itself is deployed using a dedicated physical server or an independent virtual machine instance, and does not share computing resources with other applications, thereby reducing the risk of data leakage and lateral movement attacks from the computing environment level.

[0135] Minimize privileges mechanism: At the operating system level, the AI ​​agent process runs under a dedicated low-privilege system account, which does not belong to the administrator group and only has ordinary user privileges.

[0136] The file system permissions are configured as follows: read-only access is allowed to the skill pack storage directory and its subdirectories; read and write access is allowed to the task temporary working directory; no access is allowed to other sensitive system directories.

[0137] At the database level, the database connection account used by the agent is only granted permission to execute specific stored procedures, or only has permission to query, insert, and update BOM-related data tables, and does not have permission to delete, delete tables, or modify table structures.

[0138] At the API call level, the agent holds only the minimum permission token to access specific API endpoints. The scope of the token's permissions is precisely limited to accessing only specific API interfaces related to BOM management. The token has an expiration date and needs to be rotated and updated periodically.

[0139] Operational auditability mechanism and secondary verification: All operations performed by the intelligent agent during operation are recorded in the audit log system. The audit log records include the following fields: timestamp, task ID, operation type, target object of the operation, operation details, execution result, and operator identifier.

[0140] Audit logs are stored on a separate log server, physically separated from the agent server and business database server, to prevent log tampering or deletion. Logs are backed up and archived regularly, with retention periods meeting enterprise compliance requirements.

[0141] For irreversible operations (including but not limited to deleting the BOM version, modifying master data, sending external notifications, and executing system-level commands), the system has a secondary manual confirmation mechanism. The workflow of this mechanism is as follows: When an agent determines that an irreversible operation needs to be performed, it does not execute it immediately. Instead, it pauses the current task flow and generates a secondary confirmation request. This request includes a request ID, task ID, operation type, operation summary (describing the specific content of the operation to be performed in natural language), and a description of the scope of impact (the number of data entries affected by the operation, the business systems involved, etc.).

[0142] The confirmation request is sent to users with approval permissions via the management console. Users can view the details of the confirmation request in the management console and choose to confirm or reject it. If the user does not respond within the preset timeout period, the system automatically marks the operation as timed out and logs it.

[0143] If the user confirms execution, the management console notifies the agent to continue execution via an internal callback interface (such as a REST API or message queue). The agent resumes the task from the pause point, executes the operation, and records the execution result and the confirming user ID to the audit log.

[0144] If the user refuses to execute, the agent cancels the operation, logs it, and sends feedback to the user who initiated the original task, indicating that the operation was rejected.

[0145] Through the above mechanisms, it is ensured that the final execution authority is always in the hands of the operator, and all operations are fully logged for subsequent auditing, tracing, and anomaly analysis.

[0146] 10. Implementation of Supplementary Response Mechanisms BOM (Bill of Materials) Frequent Change Response Mechanism: To address the frequent changes to the Bill of Materials (BOM), the system utilizes version snapshot management to achieve complete change tracking. The BOM version table's data structure includes the following key fields: Version ID, Product Model, BOM Type, Major Version Number, Minor Version Number, Parent Version ID, Version Creation Time, Version Activation Status, and Reason for Version Creation.

[0147] During each change operation, the system first marks the currently active version as a historical state in the database transaction and records the archive timestamp; then, it uses the current version data as a template to copy and create a new version record, generates an incrementing version number, sets the parent version ID to point to the archived version to establish a traceability chain; finally, it executes the changes on the new version data.

[0148] Through the parent version ID chain tracing structure, the system supports the following version management functions: complete viewing of the BOM structure of any historical version; comparison of differences between any two versions; and fast rollback of historical versions.

[0149] In addition, the system utilizes AI to learn patterns from accumulated historical change data. The learning process uses each BOM change record as training data, including features such as change time, material code, material category, change type, and trigger reason classification label.

[0150] During operation, the system periodically acquires external condition data (such as raw material market price index and average delivery cycle of key suppliers) and compares and analyzes these external conditions with the historical change patterns of each material in the current active BOM.

[0151] For materials that are highly similar to historical change patterns, the system marks them as high-risk materials and automatically generates a list of pre-adjustment suggestions, including material codes, predicted change types, historical reference cases, and confidence scores, which are then pushed to BOM managers for reference in advance.

[0152] Mechanism for handling non-standardized material data: The documentation for the BOM verification skills module includes a complete data cleaning pipeline process, which consists of two steps: The first step is rule cleaning. Using a preset set of text cleaning regular expressions, the material name and specification fields are automatically cleaned, including removing redundant whitespace characters, converting full-width punctuation marks to half-width characters, automatically replacing common spelling errors and synonyms, and standardizing unit symbols.

[0153] The second step is vector similarity matching. The cleaned material name and specifications are concatenated into a text string, which is then converted into a vector representation using text embedding technology. Next, a nearest neighbor search is performed in a pre-built enterprise standard material master data vector index library, and cosine similarity is calculated. If the highest similarity found exceeds a preset threshold, the material description is automatically replaced with the corresponding standard material's specification name and model, and this automatic mapping operation is recorded in the verification report. If the highest similarity is below the threshold, the material is marked as requiring manual confirmation, and the top three candidate standard material entries with the highest similarity are attached to the verification report.

[0154] This data cleaning pipeline can process all material items in the entire BOM in batches and output a standardized mapping report after processing.

[0155] Cross-departmental collaborative response mechanism: After executing a change operation, the BOM change management skills module automatically distributes change notifications through a message middleware. The notification message body uses a structured JSON format, including information fields such as change notification ID, change order number, change type, change initiator, change effective date, list of affected product models and corresponding BOM version number, and a summary of the change content.

[0156] The message push script reads the preset notification rule configuration file, matches the corresponding list of receiving departments according to the type of the current change, and then distributes the notification message to the corresponding department group or personnel through the enterprise message middleware.

[0157] Meanwhile, the system monitors cost changes in real time during the change impact assessment phase. When calculating the scope of impact, the change execution script obtains the unit standard cost data of the materials involved in the change through the cost accounting interface and automatically calculates the total cost difference before and after the change. When the calculated absolute value of the cost change exceeds a preset threshold, the system's notification level is automatically escalated to an alert. The alert notification message adds an alert identifier and an urgency level label to the regular notification, and is additionally pushed to the finance and supply chain management departments in addition to notifying the regular receiving departments. At the same time, the system automatically triggers a manual approval process through the workflow interface, requiring managers with the corresponding approval authority to review changes with high cost impacts.

[0158] XI. A Complete Example of Module Collaboration To more clearly demonstrate the data flow and collaboration between the four functional modules, the following example uses a complete scenario of a new product introduction to illustrate the collaborative working process of each module.

[0159] An electronics manufacturing company initiates a new product introduction process. The BOM administrator issues a task instruction to the intelligent agent, requiring it to generate a complete manufacturing BOM for the new product, verify it, and synchronize it to the ERP and MES systems after confirming that it is correct.

[0160] (I) BOM Creation Phase The agent identifies the BOM-creator skill through intent recognition in the ReAct inference loop. After loading the documentation, the agent sequentially calls the design data parsing script, the EBOM to MBOM conversion script, and the material code generation script. It extracts the EBOM structure from the design files obtained from the PLM system, converts it into an MBOM containing process virtual parts and auxiliary materials, and generates material codes conforming to the company's coding standards for new materials. Finally, it generates an MBOM data file, which serves as input data for subsequent modules.

[0161] (II) BOM Verification Phase Following the instructions in the subsequent suggestion fields of the BOM-creator documentation, the agent automatically invokes the BOM-validator skill after BOM creation is complete. The agent takes the path to the MBOM file generated in the previous stage as input, loads the BOM-validator documentation, and sequentially performs rule engine validation and knowledge graph error correction.

[0162] During the verification process, the following issues were discovered: The name of a certain material was inconsistent with the standardized name in the enterprise's standard material library. This was automatically corrected after matching using a knowledge graph, with a confidence level of 0.92. The packaging type field for a certain material was empty, marked as a missing required field, and required manual completion. The agent generated a structured verification report, listing the anomalies and their handling suggestions, and presented it to the BOM administrator.

[0163] (III) Change and Correction Phase After reviewing the validation report, the BOM administrator accepted the automatically corrected content and manually supplemented the missing encapsulation type information. Since these corrections involved modifying the BOM data, the agent invoked the bom-change-manager skill according to the correction trigger rules in the bom-validator documentation.

[0164] The change management skill receives the revised content and performs an impact assessment. Since the current MBOM is newly created and has no associated inventory records, in-transit purchase orders, or incomplete production work orders, the impact assessment result is zero. The agent generates a new version of the MBOM data file, writes the revised data into the new version, retains the original version as a historical version, and records change traceability information.

[0165] (iv) Cross-system synchronization phase After successful verification and correction, the BOM administrator confirms the MBOM data is correct and issues a command to the agent to synchronize it to the business systems. The agent invokes the bom-sync skill, loads its documentation, and obtains the list of target systems.

[0166] For ERP systems with standard APIs, the agent invokes the API interaction script to push MBOM data to the ERP's BOM import interface via an HTTP POST request. For older MES systems without standard APIs, the agent invokes a screen understanding synchronization script, launches a controlled browser through a browser automation framework, uses an interface element detection model to identify operation elements, and simulates manual clicks and inputs to complete the import of MBOM data onto the MES system interface.

[0167] During the synchronization process, the coding mapping script automatically translates the material codes in the MBOM into the corresponding material codes in the ERP and MES systems. After synchronization is complete, the agent performs field-level consistency checks (comparing checksums of key fields) on the BOM data in each target system with the source MBOM file. Once it confirms that the synchronization is complete and error-free, it reports the task completion information to the BOM administrator.

[0168] Thus, a complete new product BOM import process is efficiently completed through the autonomous scheduling of the intelligent agent and the coordinated cooperation of various skill modules.

[0169] Example 1

[0170] This paper takes an electronics manufacturing company as an example. This company produces printed circuit boards and processes approximately 20 to 30 BOM change requests daily, involving operations such as electronic component replacement, package specification adjustment, and alternative material switching. After adopting the method of this invention, the following effects are achieved.

[0171] The BOM change processing cycle has been shortened from an average of 2.5 working days to approximately 30 minutes. This is achieved by the BOM change management skill module, which automatically executes all tasks—such as understanding change requests, retrieving affected BOMs, querying inventory and procurement data, manually updating multiple systems, and writing and distributing change notices—through a ReAct inference loop. This transforms the previously manual process, reliant on multiple departments for manual approval, into automated parallel processing.

[0172] The material code duplication rate decreased from approximately 3.5% before implementation to below 0.1%. This effect is achieved by introducing an automatic material code generation and uniqueness verification mechanism in the BOM creation module, and an automatic review of code uniqueness rules in the BOM verification module. These two aspects, code generation and verification, jointly ensure the uniqueness of the codes.

[0173] The timeliness rate of cross-departmental change notifications increased from approximately 70% before implementation to 100%. The principle behind this achievement lies in the automatic notification closed-loop mechanism of the BOM change management skills module. After a change is executed, the system automatically matches the receiving department according to preset rules and pushes the notification in real time, eliminating omissions and delays that may occur with manual notifications.

[0174] Example 2

[0175] This paper takes an industrial equipment manufacturing company as an example. This company has a high proportion of non-standard customized products, with a single product's BOM reaching more than eight levels and over 2000 material items, resulting in complex design and production coordination. After adopting the method of this invention, the following effects were achieved.

[0176] The conversion time from EBOM to MBOM has been reduced from an average of 3 working days to approximately 2 hours. This is achieved by automating all steps in the BOM creation module, including design data parsing, automatic addition of virtual process components, automatic matching of auxiliary materials, and batch generation of material codes. This replaces a significant amount of manual work and cross-system queries by process engineers, demonstrating the efficiency gains brought about by knowledge assetization.

[0177] The BOM version management achieves 100% completeness, with all historical changes traceable to specific versions and operators. This is achieved through a version snapshot management mechanism, which automatically generates a new version with each change and maintains a complete traceability chain through the parent version ID, completely eliminating the problems of version omissions or confusion caused by traditional manual numbering methods.

[0178] Production anomalies caused by BOM data inconsistencies decreased by approximately 85% year-on-year. This achievement is due to the automatic synchronization and consistency verification mechanism of the BOM cross-system synchronization skill module, as well as the material coding semantic mapping technology, which solves the problem of inconsistent material identification between different systems and ensures real-time consistency of BOM data in PLM, ERP, and MES systems.

[0179] The above description is merely a preferred embodiment of the present invention and does not limit the scope of the patent. For those skilled in the art, different specific implementation tools and parameter configurations can be selected based on the actual IT environment and business needs of the enterprise without departing from the technical principles of the present invention. Any equivalent structural or procedural transformations made using the content of this specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of this invention.

Claims

1. A method for automated BOM management based on AI intelligent agents and skill package mechanisms, characterized in that, Includes the following steps: Step S1: Encapsulate the professional knowledge in the domain of bill of materials management into multiple skill packages for AI agents to call. Each skill package adopts a three-level progressive loading architecture, which includes: a metadata layer that is always loaded when the agent starts, a documentation layer that is loaded on demand when a task is triggered, and a resource and code layer that is called on demand during the execution phase without entering the agent's context window. Step S2: Construct a set of functional modules that include at least the following: BOM creation skill module, BOM verification skill module, BOM change management skill module, and BOM cross-system synchronization skill module; Step S3: The skill package is attached to the AI ​​agent through the skill registration framework, so that the agent can discover and call the registered skills during runtime; Step S4: Drive the AI ​​agent to perform BOM management tasks through the ReAct inference loop. The ReAct inference loop includes: scanning the metadata layer of registered skills for intent recognition and skill matching, loading the description document layer of the matched skills to obtain the operation process and rules, calling resources and code layers to perform operations, and generating BOM data files and pushing them to the target business system for result output and synchronization.

2. The BOM automated management method based on AI intelligent agent and skill package mechanism according to claim 1, characterized in that, In the three-tier progressive loading architecture of step S1: The metadata layer includes skill names and function descriptions, and the metadata layer token consumption for a single skill package is less than 100 tokens; The documentation layer contains operation guidelines, workflows, and constraint rules. The documentation layer token consumption for a single skill pack is less than 5000 tokens, and it only occupies the agent context window during task execution. The resource and code layer contains executable scripts, reference documents, BOM structure templates, and validation rule sets. This layer of resources does not enter the agent context window; it is only loaded and run by the execution environment when used.

3. The BOM automated management method based on AI intelligent agent and skill package mechanism according to claim 1, characterized in that, The execution flow of the BOM creation skill module in step S2 includes: Extract structured data from the design BOM from the CAD / PLM system; Convert the design BOM into a manufacturing view and add process virtual parts and auxiliary material information; Batch generate or verify material codes by referencing material coding rule files; Output manufacturing BOM data file.

4. The BOM automated management method based on AI intelligent agent and skill package mechanism according to claim 1, characterized in that, The verification mechanism of the BOM verification skill module in step S2 includes: Based on a preset rule set, the BOM data is automatically audited in batches to ensure the completeness of required fields, the uniqueness of material codes, and the correctness of parent and child item usage. Using knowledge graphs in the manufacturing sector, automatic error correction of material names and specifications is achieved through entity recognition and semantic matching; It also generates a structured verification report that identifies anomalies, anomaly types, and suggested corrections.

5. The BOM automated management method based on AI intelligent agent and skill package mechanism according to claim 1, characterized in that, The execution flow of the BOM change management skills module in step S2 includes: Receive change instructions and extract the change type, target material, and change parameters through natural language processing; Automatically retrieve all BOM versions containing the target material and assess the impact of changes on inventory, in-transit purchases, work-in-process orders, and costs; Perform batch modification of BOM master data and generate new version numbers while retaining the historical version traceability chain, and synchronize the update to related systems; Change notifications are automatically distributed according to preset rules to form a closed loop.

6. The BOM automated management method based on AI intelligent agent and skill package mechanism according to claim 1, characterized in that, The synchronization methods for the BOM cross-system synchronization skill module in step S2 include: For legacy systems that do not support standard APIs, intelligent screen understanding technology is used to visually recognize system interface elements and simulate manual operations to complete data reading and writing. The semantic understanding capabilities of large language models are used to identify the differentiated codes and field representations of the same material across different systems and to establish a coding mapping relationship table. It supports local deployment on enterprise servers or private cloud environments, ensuring that data does not leave the enterprise network boundaries.

7. The BOM automated management method based on AI intelligent agent and skill package mechanism according to claim 1, characterized in that, The ReAct inference loop in step S4 specifically includes: S4-1 Intent Recognition and Skill Matching: Matches the semantics of the user's task with the metadata layer information of all registered skills to filter out the most relevant skill package combinations; S4-2 Knowledge Loading: Read the contents of the description document layer of the matching skill and load it into the agent's context window to obtain the operation process, constraint rules and processing strategies; S4-3 Action Execution: Following the instructions in the documentation, invoke the resources of the matching skills and the executable scripts and resources in the code layer to perform BOM management operations; S4-4 Result Output and Synchronization: Generate a BOM data file, and after manual confirmation, call the BOM cross-system synchronization skill module to push it to the target business system.

8. The BOM automated management method based on AI intelligent agent and skill package mechanism according to any one of claims 1 to 7, characterized in that, It also includes safety protection measures: Implement network or physical layer isolation between the intelligent agent's operating environment and the production system. Deploy the intelligent agent on a dedicated server or independent virtual machine instance, and only open the network ports and access paths necessary for the task. Configure permissions to minimize the file system directories, application programming interfaces and network addresses accessible to the agent, granting only the minimum set of permissions necessary to complete the scheduled BOM management tasks; All operations performed by the intelligent agent during its operation are logged, and a second manual confirmation step is set up for irreversible operations to ensure that the final execution authority is in the hands of the operator.

9. The BOM automated management method based on AI intelligent agent and skill package mechanism according to any one of claims 1 to 7, characterized in that, It also includes at least one of the following response mechanisms: The BOM high-frequency change response mechanism uses version snapshot management to automatically generate a new version number and retain a complete historical version traceability chain for each change. It also uses AI to learn patterns from historical change data to predict BOM adjustment needs and generate pre-adjustment suggestions. The non-standard material data handling mechanism includes a data cleaning pipeline pre-set in the BOM verification skill module, which comprehensively uses rule matching and vector similarity calculation to automatically map non-standard material names and specifications to enterprise standard material master data entries. A cross-departmental collaborative response mechanism automates change approval and cross-system synchronization through the automatic notification function of the BOM change management skills module. When the cost change caused by the BOM change exceeds the preset threshold, an early warning notification is automatically generated and a manual intervention process is triggered.

10. A BOM (Bill of Materials) automated management system based on AI intelligent agents and skill package mechanisms, characterized in that, include: The skill package building module is used to encapsulate knowledge in the bill of materials management domain into multiple skill packages using a three-tier progressive loading architecture. The three tiers include a metadata layer, a documentation layer, and a resource and code layer. Functional module unit, used to manage a set of functional modules including at least a BOM creation skill module, a BOM verification skill module, a BOM change management skill module, and a BOM cross-system synchronization skill module; The skill registration and mounting module is used to mount the skill package to the AI ​​agent instance through the skill registration framework; An automated execution module is used to drive an agent to perform BOM management tasks through a ReAct inference loop, wherein the ReAct inference loop includes intent recognition and skill matching, knowledge loading, action execution, and result output and synchronization. The security protection module is used to implement runtime environment isolation, permission minimization control, and operation auditing.

Images