A Processing Route Generation Method Based on Knowledge Graph and Review Agent

By constructing a knowledge graph and a process route generation method based on review agents, the problem of insufficient reliability in process route generation in existing technologies is solved, and high-quality generation of process routes and continuous optimization of the knowledge base are achieved, forming a closed loop of process route generation, review and knowledge updating.

CN123117296APending Publication Date: 2026-11-17NORTHWESTERN POLYTECHNICAL UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611017679.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-09
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing process route generation methods lack reliability, adaptability to knowledge base, and continuous optimization capabilities in complex manufacturing scenarios. Furthermore, the feedback linkage and collaborative closed loop between various stages are imperfect, making it difficult to meet the needs of high-quality generation and dynamic optimization.

Method used

The process route generation method based on knowledge graph and review agent constructs a mechanical product manufacturing process knowledge graph, rule base and retrieval-enhanced review knowledge base, and uses review agent to conduct evidence-based review and correction of the initial process route, forming a closed loop of process route generation, review, correction and knowledge update.

Benefits of technology

It improves the reliability and interpretability of process route generation, enhances the reusability of process knowledge and adaptability to complex manufacturing scenarios, avoids the recurrence of similar problems in subsequent route generation, improves generation quality, and achieves continuous optimization of the knowledge base.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117296A_ABST
    Figure CN123117296A_ABST
Patent Text Reader

Abstract

To address the shortcomings of existing process route generation methods, such as insufficient reliability, inadequate knowledge base adaptability, and insufficient continuous optimization capabilities, which make it difficult to meet the comprehensive demands for high-quality process route generation, dynamic optimization, and continuous knowledge evolution in complex manufacturing scenarios, this invention proposes a process route generation method based on knowledge graphs and review agents. This invention generates an initial process route using a knowledge graph and rule base, then uses a retrieval-enhanced review knowledge base and a review agent to conduct evidence-based review and targeted corrections of the initial process route. The review results are then categorized and fed back to the corresponding knowledge base, forming a closed loop of process route generation, review, correction, and knowledge updating. This improves the reliability and interpretability of process route generation, enhances the reusability of process knowledge, and strengthens the adaptability and continuous optimization capabilities of the knowledge base to complex manufacturing scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a method for generating process routes. Background Technology

[0002] The process route is the core link connecting product design information and the manufacturing execution process. It is used to determine the processing methods, sequence of operations, inspection arrangements, and resource allocation schemes for the object to be processed. Traditionally, process route development relies mainly on the professional knowledge and experience of process engineers, who complete the process by consulting process standards, referring to historical cases, and combining personal experience with judgment. However, when faced with complex processing characteristics, stringent quality requirements, or limited manufacturing resources, manual development methods often fail to fully consider various constraints, easily leading to problems such as improper process settings, unreasonable sequence arrangements, or omissions in inspection steps. Furthermore, the basis for route generation and adjustment is usually scattered across various documents, making unified traceability and reuse difficult.

[0003] In recent years, the development of knowledge graphs, artificial intelligence, and retrieval-enhanced generation technologies has enabled process standards, historical routes, and expert experience to participate in process route generation in a structured or searchable form. However, to ensure the engineering usability of process routes, it is not only required that the process route generation method can generate a sequence of operations, but more importantly, each operation in the generated sequence must meet the process requirements of processing characteristics, quality inspection specifications, logical constraints between operations, and manufacturing resource capacity limitations. Furthermore, the process route generation results must undergo rationality and feasibility verification.

[0004] Research has emerged on the organization of process knowledge, generation of process routes, and evaluation of generation results, for example:

[0005] Patent application CN119378350A discloses a method for generating and allocating process routes for complex components based on a dynamic knowledge graph. This method constructs a knowledge graph by integrating historical process knowledge and dynamic production data, generates process routes using a graph recurrent neural network, and then optimizes allocation using an ant colony algorithm. This method focuses on improving the automation level of process route generation and task allocation, but its processing does not consider the review and feedback correction of the generated results. When problems such as missing processes or sequence deviations exist in the generated process routes, it is difficult to detect and correct them in a timely manner, adversely affecting the reliability of the final result.

[0006] A patent application with authorization announcement number CN121303313B discloses a knowledge-driven automatic generation method for parts processing technology. Its core lies in generating a processing method chain through a two-way matching of processing requirements and enterprise resource capabilities, and coupling a knowledge graph with a large knowledge model to provide optimization suggestions for process decisions. Simultaneously, this method also assesses the credibility of newly generated knowledge during the reasoning process, updating the knowledge graph and vector index library with knowledge increments that meet the credibility threshold. However, the credibility assessment of this method mainly serves knowledge filtering and graph updating, and cannot evaluate the effectiveness of process decision results. Furthermore, in some cases, the credibility assessment of this method may allow inaccurate knowledge to be included in the knowledge graph. Such knowledge may still be reused in subsequent retrieval and reasoning, thus adversely affecting the knowledge base and the reliability of process decisions.

[0007] Patent application CN120973923A discloses a "generation-self-examination-correction" method for aircraft process instructions guided by a large language model. This method utilizes retrieval-enhanced generation technology and the synergy between the inference model and the examination model to achieve the generation, examination, and iterative correction of process instructions. It also optimizes the inference model or supplements the retrieved knowledge based on the error risk level. While this method focuses on improving the quality of generated process instructions, it fails to further analyze and trace the root causes of errors and identify corresponding update targets. This makes it difficult to translate the review results into effective supplementation of the knowledge base, hindering the continuous reuse of feedback information.

[0008] In summary, existing technologies still suffer from problems in process knowledge organization and intelligent process planning, including insufficient reliability of generated results, lack of traceability in the review process, and inadequate adaptability and continuous optimization capabilities of the knowledge base. Although the technical solutions disclosed in the aforementioned patent applications provide corresponding processing strategies for individual aspects such as process route generation, process knowledge updating, and generation result review, the feedback linkage and collaborative closed loop between these stages remain imperfect, making it difficult to meet the comprehensive needs for high-quality generation, dynamic optimization, and continuous knowledge evolution of process routes in complex manufacturing scenarios. Summary of the Invention

[0009] Based on the above background technology, in order to solve the technical problems of insufficient reliability of the generated results, insufficient adaptability of knowledge base and continuous optimization capability of existing process route generation methods, and imperfect feedback linkage and collaborative closed loop between various links, which make it difficult to meet the comprehensive needs of high-quality generation, dynamic optimization and continuous knowledge evolution of process routes in complex manufacturing scenarios, this invention proposes a processing process route generation method based on knowledge graph and review agent.

[0010] The technical solution of this invention is:

[0011] The method for generating processing routes based on knowledge graphs and review agents is unique in that it includes the following steps:

[0012] Step 1: Construct a knowledge graph of mechanical product manufacturing processes, a rule base for mechanical product manufacturing processes, and a knowledge base for enhanced retrieval and review. The knowledge base for enhanced retrieval and review consists of original text fragments extracted from mechanical product process data that can serve as the basis for process route review, text vectors after the original text fragments are segmented and vectorized, metadata records corresponding to the original text fragments, vector indexes of the text vectors, and the association between the text vectors and metadata records. Metadata records include the document name, clause location, and version information.

[0013] Step 2: Convert the processing information of the object to be processed into standardized process input data, including processing characteristics, quality requirements, resource constraints, and the relationship between quality requirements and processing characteristics or the object to be processed as a whole;

[0014] Step 3: Generate the initial process route based on standardized process input data, mechanical product manufacturing process knowledge graph, and mechanical product manufacturing process rule base;

[0015] Step 4: Construct a review agent based on retrieval enhancement generation technology and large language model, and use the review agent to review the initial process route and generate structured review results.

[0016] The review agent comprises a review unit construction module, a retrieval query statement generation module, a review evidence generation module, a review prompt information generation module, and a large language model, used to implement the following steps:

[0017] Step 4.1: The review unit construction module divides the initial process route into multiple review units for different review objects according to four review types: process integrity, quality control, process sequence, and resource matching. The review unit includes review type, review object, route segment to be reviewed, constraint information, and review purpose.

[0018] Step 4.2: The query statement generation module generates a corresponding query statement for each review unit according to the preset query template.

[0019] Step 4.3: For each review unit, the review evidence generation module uses its corresponding search query to retrieve review evidence from the enhanced review knowledge base, and sorts the retrieved review evidence according to its relevance to the current review unit, taking the top 10% to constitute review evidence.

[0020] Step 4.3: The review prompt information generation module combines the initial process route, review unit, and review evidence into review prompt information for each review unit.

[0021] Step 4.4: The large language model reviews each review unit separately: The review prompt information corresponding to the review unit is input into the large language model to determine the process route problem. The large language model outputs the structured review results of the current review unit, including the review conclusion, route problem type, problem location, problem cause, source of evidence and suggested handling method.

[0022] Step 5: Generate the target process route based on the structured review results: When the review conclusion indicates that there are no problems with the initial process route, the initial process route is used as the target process route; when the review conclusion indicates that there are problems with the initial process route, the corresponding route segment in the initial process route is modified according to the location of the problem and the suggested handling method to obtain the target process route.

[0023] Step 6: Complete knowledge based on structured review results: Classify the causes and sources of evidence in the structured review results, and convert reusable rule issues, graph knowledge issues and case issues into candidate options for updating the process rule base, candidate options for completing the process knowledge graph, and candidate options for updating the process case, respectively. After manual confirmation, write the candidate options into the corresponding mechanical product manufacturing process knowledge graph, mechanical product manufacturing process rule base or retrieval enhancement review knowledge base.

[0024] Furthermore, in step 1:

[0025] The method for constructing a knowledge graph of mechanical product manufacturing processes is as follows: extract text from mechanical manufacturing process standards, historical process cases of mechanical products, and mechanical product process experience documents; perform segmentation and terminology standardization on the extracted text; extract process object entities, process activity entities, and resource detection entities from the standardized text; identify the control relationships, resource matching relationships, and process sequence relationships between entities; and convert entities into nodes and relationships into edges to form a knowledge graph of mechanical product manufacturing processes.

[0026] The method for constructing the mechanical product manufacturing process rule base is as follows: extract process constraints from mechanical manufacturing process standards, historical process cases of mechanical products, and mechanical product process experience documents, and convert the process constraints into different rule items containing applicable objects, triggering conditions, and constraint relationships. These rule items constitute the mechanical product manufacturing process rule base.

[0027] The method for constructing the retrieval-enhanced review knowledge base is as follows: Mechanical manufacturing process standards, audited manufacturing site issues and handling records, and historical process cases are used as review knowledge sources. Original text fragments that can serve as the basis for process route review are extracted from these sources, and metadata records are established for each fragment. The original text fragments are then segmented according to semantic integrity, and the resulting text is vectorized to obtain text vectors. These text vectors are associated with and stored in relation to their corresponding metadata records, and a vector index is established based on the text vectors. Finally, the original text fragments, text vectors, metadata records, vector indexes, and the association between text vectors and metadata records together constitute the retrieval-enhanced review knowledge base. The text vectors and vector indexes are stored in the Milvus vector database.

[0028] Furthermore, step 2 specifically involves:

[0029] Step 2.1: Extract and standardize the processing information of the object to be processed, and then collect the standardized information into the feature description field, quality labeling field, and resource restriction field to obtain structured processing information.

[0030] Step 2.2: Based on the pre-built process identification dictionary and identification rules, identify processing features, quality requirements and resource constraints from the structured processing information obtained in Step 2.1, and establish the correlation between quality requirements and processing features or the whole object to be processed;

[0031] Step 2.3 involves encapsulating the identified processing features, quality requirements, resource constraints, and the relationship between quality requirements and processing features or the object to be processed as a whole into standardized process input data.

[0032] Furthermore, the process identification dictionary in step 2.2 is used to extract processing feature information, quality requirement information, and manufacturing resource information from the processing information; the process identification dictionary is constructed based on the mechanical product process data and includes a processing feature dictionary, a quality requirement dictionary, and a manufacturing resource dictionary; the identification rules are used to associate the identified processing feature information with the quality requirement information; the identification rules are constructed based on the mechanical product process data.

[0033] Furthermore, step 3 specifically involves:

[0034] Step 3.1: Using the processing features and quality requirements in the standardized process input data obtained in Step 2 as search conditions, search for matching initial candidate processes in the mechanical product manufacturing process knowledge graph constructed in Step 1, and filter out candidate processes in combination with resource constraints.

[0035] Step 3.2: Associate the selected candidate processes with their corresponding processing characteristics, quality requirements, and resource constraints to form a set of candidate processes;

[0036] Step 3.3: Call the mechanical product manufacturing process rule library established in Step 1 to establish constraint relationships for each candidate process in the candidate process set, including sequential relationships, dependency relationships and prohibition relationships, so as to organize the scattered candidate processes into a process directed graph;

[0037] Step 3.4: Perform topological sorting or constraint solving on the directed graph of the process to obtain the initial process route.

[0038] Furthermore, the preset query templates in step 4.2 include a processing integrity query template, a quality control query template, a process sequence query template, and a resource matching query template, each corresponding to one of the four review units; the specific details of each query template are as follows:

[0039] Process integrity query template: Material / part type + Processing features + Quality requirements + Current existing processes + Processing processes and process integrity requirements;

[0040] Quality control query template: Processing characteristics + Quality requirements + Existing inspection / control procedures + Quality control and inspection procedure requirements;

[0041] Work process sequence query template: Current route segment + Relationship between adjacent work processes + Work process sequence requirements;

[0042] Resource matching query template: Process to be reviewed + Quality requirements + Matched resources + Resource capabilities meet requirements.

[0043] Furthermore, the large language model in step 4.4 adopts Qwen2.5-72B-Instruct.

[0044] Furthermore, in step 6: the candidate options for updating the process rule base include the applicable objects, triggering conditions, constraint relationships, and evidence sources; the candidate options for completing the process knowledge graph include entities, relationships, and evidence sources; and the candidate options for updating the process case include process route adjustment records and the target process route generated after the process route adjustment.

[0045] The beneficial effects of this invention are:

[0046] 1. This invention generates an initial process route through a knowledge graph and rule base, and uses a retrieval-enhanced review knowledge base and a review agent to conduct evidence-based review and targeted correction of the initial process route. The review results are then categorized and fed back to the corresponding knowledge base (knowledge graph, rule base, or retrieval-enhanced review knowledge base), forming a closed loop of process route generation, review, correction, and knowledge updating. This improves the reliability and interpretability of process route generation, enhances the reusability of process knowledge, and strengthens the adaptability and continuous optimization capabilities of the knowledge base to complex manufacturing scenarios.

[0047] 2. This invention feeds back the rule gaps, knowledge relationship errors, and insufficient case studies found during the review process to the corresponding knowledge base. On the one hand, this avoids the recurrence of similar problems in subsequent route generation, thus improving the quality of process route generation. On the other hand, it transforms the review results into supplements to the knowledge base, enabling the feedback information to be continuously reused.

[0048] 3. The processing route generated by this invention can refer to a processing route that covers "casting, forging, welding, heat treatment and assembly processes", or it can refer to a processing route that only considers cutting processes (turning, milling, planing, grinding, drilling) and some forming processes (stamping). Attached Figure Description

[0049] Figure 1 This is a flowchart of the method of the present invention.

[0050] Figure 2 This is a flowchart illustrating the construction process of the process knowledge graph, process rule base, and retrieval-enhanced review knowledge base in this invention.

[0051] Figure 3 This is a flowchart of the initial process route generation, review, and correction in this invention.

[0052] Figure 4 This is a flowchart of the review result classification feedback and knowledge completion process in this invention.

[0053] Figure 5 This is a schematic diagram of the structure of a titanium alloy mounting bracket for a certain type of aero-engine in an embodiment of the present invention. Detailed Implementation

[0054] The present invention will now be described in detail with reference to the accompanying drawings.

[0055] Reference Figure 1-4 The processing route generation method based on knowledge graph and review agent proposed in this invention includes the following steps:

[0056] Step 1: Construct a knowledge graph of mechanical product manufacturing processes, a rule base for mechanical product manufacturing processes, and a knowledge base for enhanced retrieval and review.

[0057] This step focuses on building the knowledge base to support process route generation and review. Addressing the issue that mechanical product manufacturing process knowledge is scattered across mechanical manufacturing process standards, historical process cases, process experience documents, and audited records of on-site problems and their handling, making it difficult to directly participate in rule-based reasoning and evidence retrieval, this invention employs a combination of entity relation extraction, rule-based representation, and vectorized indexing to organize the scattered process knowledge into a mechanical product manufacturing process knowledge graph, a mechanical product manufacturing process rule base, and a retrieval-enhanced review knowledge base.

[0058] Reference Figure 2 The method for constructing a knowledge graph of mechanical product manufacturing processes is as follows: extract text from mechanical manufacturing process standards, historical process cases of mechanical products, and mechanical product process experience documents; perform segmentation and terminology standardization on the extracted text; extract process object entities, process activity entities, and resource detection entities from the standardized text; identify control relationships, resource matching relationships, and process sequence relationships between entities; convert entities into nodes and relationships into edges to form a knowledge graph of mechanical product manufacturing processes.

[0059] Reference Figure 2 The method for constructing the mechanical product manufacturing process rule base is as follows: extract process constraints from mechanical manufacturing process standards, historical process cases of mechanical products, and mechanical product process experience documents, and convert the process constraints into different rule items containing applicable objects, triggering conditions, and constraint relationships. These rule items constitute the process rule base.

[0060] Reference Figure 2 The method for constructing the enhanced retrieval review knowledge base is as follows: Mechanical manufacturing process standards, audited manufacturing site issues and handling records, and historical process cases are used as review knowledge sources. Original text fragments that can serve as the basis for process route review are extracted from each review knowledge source, and metadata records are established for each original text fragment. These metadata records include the document name, clause location, and version information. Subsequently, the original text fragments are segmented according to semantic integrity. The resulting text is vectorized to obtain text vectors, which are then associated and stored with the corresponding metadata records. A vector index is established based on the text vectors. The original text fragments, text vectors, metadata records, vector indexes, and the association between text vectors and metadata records together constitute the enhanced retrieval review knowledge base. The text vectors and vector indexes are stored in a Milvus vector database. When new or revised information is added to the review knowledge sources, incremental vectorization and vector index updates can be used to update the enhanced retrieval review knowledge base.

[0061] Step 2: Convert the processing information of the object to be processed into standardized process input data;

[0062] This step addresses the issues of inconsistent terminology, inconsistent field formats, and unclear correlation between quality requirements and processing characteristics in the processing information of the objects to be processed (including part engineering drawings, processing task sheets, and manufacturing resource lists). It converts the processing information of the objects to be processed into standardized process input data, which will then serve as standardized input for graph retrieval and rule reasoning.

[0063] Reference Figure 3 The specific method is as follows:

[0064] Step 2.1: Extract and standardize the processing information of the object to be processed. Collect the extracted and standardized information into the feature description field, quality labeling field and resource restriction field respectively to obtain structured processing information.

[0065] Step 2.2: Based on the pre-built process identification dictionary (constructed according to the mechanical product process data, including a processing feature dictionary, a quality requirement dictionary, and a manufacturing resource dictionary) and identification rules (constructed according to the mechanical product process data), processing features, quality requirements, and resource constraints are identified from the structured processing information obtained in Step 2.1, and the association between quality requirements and corresponding processing features or the entire object to be processed is established; resource constraints serve as input conditions for subsequent candidate process screening and resource matching review; the aforementioned process identification dictionary is used to extract processing feature information, quality requirement information, and manufacturing resource information from the processing information; the aforementioned identification rules are used to associate the processing feature information identified from the structured processing information based on the process identification dictionary with the quality requirement information. For example, if “Φ8H7”, a positioning hole, and a reamer are identified from the structured processing information based on the process identification dictionary, then “Φ8H7” is associated with the positioning hole according to the identification rules.

[0066] Step 2.3 involves encapsulating the identified processing features, quality requirements, resource constraints, and the relationship between quality requirements and processing features or the object to be processed as a whole into standardized process input data.

[0067] Step 3: Generate the initial process route based on standardized process input data, mechanical product manufacturing process knowledge graph, and mechanical product manufacturing process rule base;

[0068] This step selects candidate processes from the mechanical product manufacturing process knowledge graph based on standardized process input data, and generates an initial process route based on the constraint relationships between the candidate processes determined by the mechanical product manufacturing process rule base.

[0069] Reference Figure 3 The specific method is as follows:

[0070] Step 3.1: Using the processing features and quality requirements in the standardized process input data obtained in Step 2 as search conditions, search for matching initial candidate processes in the mechanical product manufacturing process knowledge graph constructed in Step 1, and filter out candidate processes in combination with resource constraints.

[0071] Step 3.2: Associate the selected candidate processes with their corresponding processing characteristics, quality requirements, and resource constraints to form a set of candidate processes;

[0072] Step 3.3: Call the mechanical product manufacturing process rule library established in Step 1 to establish constraint relationships for each candidate process in the candidate process set, including sequential relationships, dependency relationships and prohibition relationships, so as to organize the scattered candidate processes into a process directed graph;

[0073] Step 3.4: Perform topological sorting or constraint solving on the directed graph of the process to obtain the initial process route.

[0074] Step 4: Review the initial process route based on retrieval enhancement generation technology and large language model, and generate structured review results;

[0075] Reference Figure 3 The specific method is as follows:

[0076] Step 4.1: Read the process nodes in the initial process route, as well as the processing characteristics, quality requirements, resource constraints, and constraints between adjacent processes associated with each process node. Based on the read information, the initial process route is divided into multiple review units. A review unit is the smallest review task formed for a specific review object in the initial process route. It includes at least a review type, a review object, a route segment to be reviewed, constraint information related to the review type (such as quality requirements, process sequence constraints, and resource constraints), and a review purpose. In this invention, review types include at least four categories: processing integrity, quality control, process sequence, and resource matching. For ease of understanding, an example review unit is given here: Process Integrity Review Unit, including: Review Type: Process Integrity; Review Object: Positioning Hole; Quality Requirement: Φ8H7; Route Segment to be Reviewed: Drilling Positioning Hole, Reaming Positioning Hole; Review Purpose: For the positioning hole Φ8H7, review whether the "Drilling Positioning Hole, Reaming Positioning Hole" process in the current route is complete.

[0077] Step 4.2: Based on the type of the review unit, generate the corresponding search query statement for each review unit according to the preset query template;

[0078] The preset query templates include at least four types: processing integrity query template, quality control query template, process sequence query template, and resource matching query template, each corresponding to one of the four review units. The specific details of each query template are as follows:

[0079] Process integrity query template: Material / part type + Processing features + Quality requirements + Current existing processes + "Processing processes and process integrity requirements";

[0080] Quality control query template: Processing characteristics + Quality requirements + Existing inspection / control procedures + "Quality control and inspection procedure requirements";

[0081] Work process sequence query template: Current route segment + Relationship between adjacent work processes + "Work process sequence requirements";

[0082] Resource matching query template: Process to be reviewed + Quality requirements + Matched resources + "Resource capabilities meet requirements".

[0083] Step 4.3: For each review unit, use the corresponding search query generated in Step 4.2 to call the search-enhanced review knowledge base built in Step 1 to search for mechanical manufacturing process standards, historical process cases, and verified manufacturing site problems and handling records. Sort the mechanical manufacturing process standards, historical process cases, and verified manufacturing site problems and handling records according to their relevance to the review unit, and take the top 10% to constitute the review evidence for that review unit.

[0084] Step 4.4: For each review unit, combine the review unit, initial process route, and review evidence to form review prompt information, and input it into a large language model (such as Deepseek, Qwen2.5-72B-Instruct, etc.) for problem judgment. The large language model outputs the structured review result of the review unit. The structured review result includes the review conclusion, route problem type, problem location, problem cause, evidence source, and suggested handling method.

[0085] Steps 4.1-4.4 above can be implemented by constructing a review agent. This review agent includes a review unit construction module, a retrieval query statement generation module, a review evidence generation module, a review prompt information generation module, and a large language model (as the reasoning core), which are used to execute steps 4.1-4.4 above, respectively.

[0086] Step 5: Generate the target process route based on the structured review results;

[0087] Reference Figure 3 The system reads the review conclusions, route problem types, problem locations, and suggested handling methods from the structured review results. When the review conclusions indicate that the initial process route has no problems, the initial process route is output as the target process route. When the review conclusions indicate that the initial process route has problems, the corresponding route segments in the initial process route are modified according to the problem location and suggested handling methods. This includes process addition, sequence adjustment, testing process addition, and / or resource replacement. The modified process route is output as the target process route. At the same time, the corresponding review evidence, problem location, and route adjustment records are recorded.

[0088] Step 6: Provide categorized feedback and knowledge supplementation based on the structured review results;

[0089] This step categorizes the causes and sources of evidence in the structured review results, and converts reusable rule issues, graph knowledge issues, and case issues into candidate options for updating the process rule base, candidate options for completing the process knowledge graph, and candidate options for updating process cases, respectively. After manual confirmation, the corresponding candidate options are written into the corresponding knowledge base to form a feedback loop after the target process route is generated.

[0090] Reference Figure 4 The specific method is as follows:

[0091] Step 6.1: For the causes of problems and sources of evidence in the structured review results, determine the categories according to the preset classification rules, and generate corresponding knowledge update candidates:

[0092] ① When the cause of the problem is the absence or error of general process constraints, generate candidate rules for updating the rule base;

[0093] The review revealed that the initial process route generated using the mechanical product manufacturing process knowledge graph and the mechanical product manufacturing process rule base contained missing or incorrect steps. This indicates that the mechanical product manufacturing process rule base may lack corresponding process rules or contain incorrect rules. For example, the structured review results might show: Problem Cause: The initial process route lacks the process rule "H7 holes should have a reaming or precision boring process"; Recommended Solution: Add a reaming process after drilling; Evidence Source: xx standard clause or xx historical case. In this case, the problem cause and recommended solution can be converted into a rule base update candidate: When the machining feature is a hole and the accuracy level is H7, a reaming or precision boring process should be set after drilling.

[0094] ② When the problem is caused by missing or incorrect entities or relationships, generate candidate options for completing the process knowledge graph;

[0095] The review revealed missing entities or relationships in the initial process route, indicating an incomplete knowledge graph of the mechanical product manufacturing process. For example, the structured review results showed: Cause of the problem: The initial process route lacks a corresponding positional accuracy detection step. This reflects a missing information in the mechanical product manufacturing process knowledge graph: "The coordinate measuring machine (CMM) can detect the positional accuracy of the positioning hole." Recommended solution: Add a "CMM detects the positional accuracy of the positioning hole" step after reaming the positioning hole. Source of evidence: A clause in a testing specification or historical testing process case. In this case, the cause of the problem, recommended solution, and source of evidence are converted into candidate options for completing the process knowledge graph: Complete entities: Coordinate measuring machine, positioning hole positional accuracy; Complete relationship: The coordinate measuring machine "can detect" the positioning hole positional accuracy.

[0096] ③ When the problem is caused by insufficient historical cases or special process requirements of the current object to be processed, generate process case update candidates;

[0097] The review revealed issues with the initial process route requiring adjustment. However, there were few relevant historical cases, no records of such adjustments, or the current processing object was unique. For example, the structured review results showed: Problem Cause: The positioning hole of the current titanium alloy mounting bracket is located on the deep cavity sidewall. After conventional drilling and reaming, there is still a risk of chip removal difficulties or unstable hole wall quality. This problem is caused by the special structure of the current processing object. Recommended Solution: Add a pre-drilling guide hole process before drilling the positioning hole, or add a chip removal / deburring auxiliary process before reaming. Evidence Source: Reviewed manufacturing site problem and handling records, or verified historical process cases of similar parts. In this case, only the special structure of the current processing object and the adjusted route segment are considered as process case update candidates: Process Case: Processing Case of Positioning Hole on Deep Cavity Sidewall of Titanium Alloy Mounting Bracket; Problem Scenario: Positioning hole on deep cavity sidewall, difficult chip removal, unstable hole wall quality; Original Route Segment: Drill positioning hole - Ream positioning hole; Adjusted Route Segment: Pre-drill guide hole - Drill positioning hole - Chip removal / deburring - Ream positioning hole.

[0098] Step 6.2: The candidates generated in Step 6.1 are reviewed manually. Based on professional experience and the source of review evidence, it is determined whether they are reasonable and whether they need to be updated. If so, the candidates that pass the review and need to be updated are written into the mechanical product manufacturing process rule base. The candidates that pass the review and need to be updated process knowledge graph completion are written into the mechanical product manufacturing process knowledge graph. The candidates that pass the review and need to be updated process case are written into the retrieval enhancement review knowledge base as verified historical process cases for reuse in subsequent reviews.

[0099] Example

[0100] This embodiment is an example of generating the machining process route for a titanium alloy mounting bracket of a certain type of aero-engine. (Refer to...) Figure 5 The titanium alloy mounting bracket includes machining features such as a reference plane 1, a weight-reducing groove 3, a positioning hole 4, and bolt mounting holes 2. The surface roughness requirement for the reference plane 1 is Ra0.8, the size requirement for the positioning hole 4 is Φ8H7, and the positional accuracy requirement for the positioning hole 4 relative to the reference plane 1 is 0.03. Available machining and inspection resources include a five-axis machining center, drilling tools, milling tools, reamers, special fixtures, and a coordinate measuring machine. Process engineers aim to generate a target process route that meets the machining features, quality requirements, and resource constraints.

[0101] This embodiment follows the process route for generating the titanium alloy mounting base according to the following steps:

[0102] Step 1: Based on the collected basic process data, construct a knowledge graph of mechanical product manufacturing processes, a rule base for mechanical product manufacturing processes, and a knowledge base for enhanced review and retrieval.

[0103] ① Construction of a knowledge graph of mechanical product manufacturing processes:

[0104] By using document parsing interfaces and optical character recognition, text (including main text, tables, and process records) is extracted from process standards (such as GB / T 41923.5—2022 "Three-dimensional process design of mechanical products Part 5: detailed design"), historical process cases, and process experience documents. The extracted text is divided into process text fragments according to chapter titles, process numbers, and semantic boundaries. The process text fragments are then standardized by synonym expression according to the terminology provisions in relevant process standards (such as GB / T 4863—2008 "Basic Terminology of Mechanical Manufacturing Processes") to obtain standardized process text fragments.

[0105] The BERT-CRF entity recognition model is used to extract process object entities (processing features, quality requirements), process activity entities (such as process methods), and resource detection entities (such as equipment, cutting tools, and detection devices) from normalized process text fragments.

[0106] The BERT relationship classification model is used to identify and extract the control relationships, resource matching relationships, and process sequence relationships between entities;

[0107] Entities are written as nodes and relationships as edges into the Neo4j graph database to form a knowledge graph of mechanical product manufacturing processes.

[0108] ② Construction of a mechanical product manufacturing process rule base:

[0109] Statements with constraint meanings are extracted from mechanical manufacturing process specifications (machining process regulations, hole machining process specifications, inspection operation instructions, etc.), historical process cases of mechanical products, and expert experience, and organized into rule items (such as "when the machining feature is a hole and the hole diameter accuracy is H7, drilling precedes reaming") and stored in the mechanical product manufacturing process rule library for subsequent generation of directed process graphs.

[0110] ③ Construct a retrieval-enhanced review knowledge base:

[0111] Texts that can be used as review basis in mechanical manufacturing process standards, historical process cases, and audited manufacturing site issues and handling records are divided into evidence fragments. Sentence vector model is used to convert the evidence fragments into evidence fragment vectors and write them into a vector database. An evidence fragment vector index is also established to form a retrieval-enhanced review knowledge base.

[0112] Step 2: Convert the machining information of the titanium alloy mounting bracket into standardized process input data.

[0113] Step 2.1: Input the machining information of the titanium alloy mounting bracket (such as "part engineering drawing", "machining task sheet" and "manufacturing resource list") and use Qwen2.5-VL-72B-Instruct multimodal large model to parse the machining information. Write the parsed content into the feature description field, quality annotation field and resource constraint field respectively, and perform unit unification and format cleaning (remove redundant spaces and duplicate characters, and correct character errors generated during drawing recognition) to obtain structured machining information. Specifically, the geometric / structural features extracted from the part engineering drawing are written into the feature description field, including "datum plane, weight reduction groove, positioning hole, bolt mounting hole"; the annotation text extracted from the machining task sheet is written into the quality annotation field, including "datum plane surface roughness requirement is Ra0.8, positioning hole size requirement is Φ8H7, positioning hole position accuracy requirement relative to datum plane is 0.03"; and the resource information extracted from the manufacturing resource list is written into the resource constraint field, including "five-axis machining center, drilling tool, reamer, special fixture and three-coordinate measuring machine".

[0114] Step 2.2 employs a dictionary-based and rule-based matching algorithm to identify machining features, quality requirements, and resource constraints from the structured machining information: Machining features “datum plane, weight reduction groove, positioning hole, bolt mounting hole” are identified from the feature description field; quality requirements “surface roughness requirement Ra0.8, size requirement Φ8H7, positional accuracy requirement 0.03” are identified from the quality annotation field; and a correlation between machining features and quality requirements is established based on the identification rules: “datum plane - surface roughness Ra0.8”, “positioning hole - size requirement Φ8H7”, “positioning hole - positional accuracy requirement relative to the datum plane 0.03”; resource constraints “five-axis machining center, drilling tool, reamer, special fixture, and coordinate measuring machine” are identified from the resource restriction field.

[0115] Step 2.3: Organize the identification results into standardized process input data, including: machining features as "datum plane, weight reduction groove, positioning hole, bolt mounting hole"; quality requirements as "datum plane surface roughness Ra0.8, positioning hole size requirement Φ8H7, positioning hole position accuracy requirement relative to datum plane 0.03"; the correspondence between machining features and quality requirements as "datum plane - surface roughness Ra0.8", "positioning hole - size requirement Φ8H7", "positioning hole - position accuracy requirement relative to datum plane 0.03"; resource constraints as "five-axis machining center, drilling tools, reamers, special fixtures and three-coordinate measuring machine".

[0116] Step 3: Generate the initial process route based on the standardized process input data, the mechanical product manufacturing process knowledge graph, and the mechanical product manufacturing process rule base.

[0117] Step 3.1: Using the machining features and quality requirements in the standardized process input data as query conditions, the mechanical product manufacturing process knowledge graph constructed in Step 1 is used to retrieve suitable candidate processes through graph query statements. Using "positioning hole - H7 precision" as the query condition, the mechanical product manufacturing process knowledge graph returns initial candidate processes such as drilling, reaming, boring, or precision boring. Then, considering the current resource constraints, candidate processes that cannot be performed by a five-axis machining center, drilling tools, reamers, special fixtures, and coordinate measuring machine are eliminated, and candidate processes that meet the resource constraints are selected. In this embodiment, for the datum plane and its surface roughness requirement Ra0.8, the processes of rough milling and finish milling of the datum plane are retrieved; for the weight reduction groove, the processes of rough milling and finish milling of the weight reduction groove are retrieved; for the positioning hole and its size requirement Φ8H7, the processes of drilling, reaming, boring, and precision boring of the positioning hole are retrieved; for the bolt mounting hole, the process of drilling the bolt mounting hole is retrieved. Subsequently, the process was screened by considering the constraints of the five-axis machining center, drilling tools, reamers, special fixtures and coordinate measuring machines. Since the manufacturing resource list did not include reaming tools and precision boring tools, the processes of reaming positioning holes and precision boring positioning holes were eliminated, resulting in the following candidate processes: rough milling of the reference plane, precision milling of the reference plane, rough milling of the weight reduction groove, precision milling of the weight reduction groove, drilling of positioning holes, reaming of positioning holes and drilling of bolt mounting holes.

[0118] Step 3.2: Associate the retained candidate operations with their corresponding machining features, quality requirements, and resource combinations to generate a sortable set of candidate operations: combine "rough milling" with "datum plane" to form "rough milling datum plane", and combine "drilling", "reaming", and "locating hole" to form "drilling locating hole" and "reaming locating hole". Each candidate operation also retains its corresponding machining features, quality requirements, and resource constraints for subsequent establishment of operation constraint relationships.

[0119] Ultimately, the generated candidate process set includes:

[0120] Rough milling of the reference plane, the corresponding machining feature is the reference plane, the requirement is to remove the reference plane allowance and retain the finish milling allowance, the resources used are a five-axis machining center, milling tools and special fixtures;

[0121] The reference plane is precision milled. The corresponding machining feature is the reference plane. The requirements are to achieve the final size of the reference plane and a surface roughness Ra0.8. The resources used are a five-axis machining center, precision milling tools and special fixtures.

[0122] Rough milling of the weight reduction groove, the corresponding machining feature is the weight reduction groove, requires removing the main allowance of the weight reduction groove and retaining the allowance of the finish milling, the resources used are a five-axis machining center, milling cutters and special fixtures;

[0123] The precision milling of the weight-reducing groove corresponds to the machining feature of the weight-reducing groove. The weight-reducing groove is formed and the resources used are a five-axis machining center, precision milling tools and special fixtures.

[0124] Drilling positioning holes, the corresponding machining feature is positioning holes, which requires forming the bottom hole before reaming. The resources used are a five-axis machining center, drilling tools and special fixtures.

[0125] The reamer is a locating hole, and the corresponding machining feature is a locating hole. The requirement is to achieve the locating hole size of Φ8H7. The resources used are a five-axis machining center, a reamer, and a special fixture.

[0126] Drilling bolt mounting holes, the corresponding machining feature is bolt mounting holes, the stage requirement is to form bolt mounting holes, and the resources used are a five-axis machining center, drilling tools and special fixtures.

[0127] Step 3.3: Based on the rule items matched in the mechanical product manufacturing process rule library constructed in Step 1, add constraint edges (i.e., directed edges) between candidate processes in the candidate process set: According to the rule that roughing precedes finishing, add a directed edge from the rough milling reference plane to the finishing milling reference plane, and add a directed edge from the rough milling weight-reducing groove to the finishing milling weight-reducing groove; according to the hole machining accuracy rule, add a directed edge from drilling the positioning hole to reaming the positioning hole; according to the reference-first rule, add a directed edge from the finishing milling reference plane to the rough milling weight-reducing groove; according to the rule that large allowance removal machining precedes hole machining, add a directed edge from the finishing milling weight-reducing groove to drilling the positioning hole, and a directed edge from the finishing milling weight-reducing groove to drilling the bolt mounting hole. After adding the directed edges, the process directed diagram is obtained. After performing topological sorting on the obtained directed graph of processes, the initial process route is generated as follows: rough milling of the reference plane → finish milling of the reference plane → rough milling of the weight reduction groove → finish milling of the weight reduction groove → drilling of the positioning hole → reaming of the positioning hole → drilling of the bolt mounting hole.

[0128] Step 4: Review the initial process route to obtain structured review results.

[0129] This embodiment employs a review agent to perform enhanced review of the initial process route. This review agent uses the Qwen2.5-72B-Instruct large language model as its inference core and is configured with review unit construction modules, retrieval query statement generation modules, review evidence generation modules, and review prompt information generation modules. Its method for reviewing the initial process route is as follows:

[0130] Step 4.1: The review unit construction module in the review agent reads the process nodes in the initial process route, as well as the processing characteristics, quality requirements, resource constraints, and constraints between adjacent processes associated with each process node. It then constructs review units according to four review types: processing integrity, quality control, process sequence, and resource matching. Due to space limitations, only one review object is selected for each type of review unit as an example here.

[0131] 1) Construct a positioning hole machining integrity review unit for positioning holes, consisting of "positioning hole - Φ8H7 - drilling / reaming";

[0132] Review type: Process integrity;

[0133] Review object: positioning holes;

[0134] The route segment to be reviewed: drilling positioning holes, reaming positioning holes;

[0135] Constraint information: The machining feature is a positioning hole, and the quality requirement is Φ8H7;

[0136] Review objective: To determine whether the current route segment contains the necessary processes to meet the machining requirements of the Φ8H7 positioning hole.

[0137] 2) Construct a quality control review unit for positioning holes, namely "positioning hole - positional accuracy requirement of 0.03 relative to the reference plane - drilling and reaming".

[0138] Review type: Quality control;

[0139] Review object: Position accuracy of positioning holes;

[0140] The route segment to be reviewed: drilling positioning holes, reaming positioning holes;

[0141] Constraint information: The positional tolerance of the positioning hole relative to the reference plane is 0.03.

[0142] Review objective: To determine whether the current route requires a positioning hole position measurement procedure.

[0143] 3) Construct a review unit for the machining sequence of the datum plane, consisting of "rough milling of the datum plane - finish milling of the datum plane";

[0144] Review type: Process sequence;

[0145] Review subject: Machining route segment on the reference plane;

[0146] The route segment to be reviewed: rough milling of the reference plane, finish milling of the reference plane;

[0147] Constraint: Roughing should precede finishing;

[0148] Purpose of the review: To determine whether the roughing and finishing milling sequence of the reference plane is reasonable.

[0149] 4) Construct a resource matching and review unit for reaming operations, consisting of "reaming positioning hole - five-axis machining center - reamer";

[0150] Review type: Resource matching;

[0151] Review subject: The process of reaming positioning holes;

[0152] Route segment to be reviewed: hinge positioning hole;

[0153] Constraint information: Locating hole Φ8H7, positional tolerance 0.03, matching resources are five-axis machining center, reamer and special fixture;

[0154] The purpose of the review is to determine whether the available resources, such as the matched equipment, cutting tools, and fixtures, are sufficient to reliably meet the quality requirements of the process.

[0155] Step 4.2: The query statement generation module in the review agent generates query statements for each review unit according to the type of review unit and a preset query template. For example, for the positioning hole machining integrity review unit, the query statement "Necessary machining operations and operation integrity requirements for the Φ8H7 positioning hole of the titanium alloy mounting bracket when the current route includes drilling and reaming positioning holes" is generated; for the positioning hole quality control review unit, the query statement "Quality control and inspection operation setting requirements for the positioning hole relative to the datum plane position accuracy of 0.03 when the current route includes drilling and reaming positioning holes" is generated; for the datum plane machining operation sequence review unit, the query statement "Opposition sequence requirements for rough milling and finish milling of the datum plane" is generated; and for the reaming positioning hole resource matching review unit, the query statement "Resource capability matching between the five-axis machining center, reamer, and special fixture used for reaming positioning holes and the dimensional accuracy and position accuracy of the Φ8H7 positioning hole of 0.03" is generated.

[0156] Step 4.3: For each review unit, the review evidence generation module in the review agent calls the retrieval enhancement review knowledge base and uses the BGE-M3 text vector model to convert the retrieval query statement of the review unit into a query vector. Based on the query vector, relevant original text fragments are retrieved from the retrieval enhancement review knowledge base in the Milvus vector database according to cosine similarity as evidence fragments. The evidence fragments are sorted according to similarity scores, and the top 10% of the evidence fragments, along with their document names, clause positions, and version information, are determined as the review evidence for that review unit.

[0157] Here are a few examples of review evidence:

[0158] 1) Machining integrity, machining requirements for Φ8H7 positioning holes

[0159] Excerpt from the original text of "Drilling Process Code": Article 14: When drilling holes with high precision, rough reaming and fine reaming should be distinguished. Floating chucks should be used as much as possible when reaming holes.

[0160] 2) Quality control: the positional accuracy of the positioning hole relative to the reference plane is 0.03.

[0161] Excerpt from the original text "Method for Inspecting the Position of Threaded Locating Holes Using a Coordinate Measuring Tool": A factory produces gear parts with four threaded locating holes, requiring their positional accuracy to be controlled within 0.02 mm. Coordinate measuring machine (CMM) technology was used to measure the positional accuracy of the four threaded locating holes on the gear parts. Through processing the measurement data, the positional accuracy of each threaded locating hole was obtained.

[0162] 3) Operation sequence: the order of operations for rough milling the reference plane and finish milling the reference plane.

[0163] Excerpt from the "General Process Rules for Cutting Machining": Article 5.6: The chamfers, roundings, and groove depths of rough machining should be increased or deepened according to the finishing allowance to ensure that the design requirements are met after finishing.

[0164] 4) Resource matching: The matching of existing processing equipment with the resource capabilities required for machining positioning holes.

[0165] Excerpt from a historical process case: A housing connector has multiple assembly positioning holes. The drawings specify high requirements for the hole diameter and positional accuracy of the hole system. This case uses a machining center with a special fixture to complete the part positioning and clamping. After drilling pre-holes, a reamer is used for reaming. Finally, the hole diameter and positional accuracy of the positioning holes are inspected.

[0166] Step 4.4: For each review unit, the review prompt information generation module in the review agent combines the initial process route, review unit and its review evidence according to the preset prompt template to form the review prompt information corresponding to each review unit. The review prompt information corresponding to each review unit is input into the Qwen2.5-72B-Instruct large language model for problem judgment. The large language model outputs the structured review results corresponding to each review unit according to the preset structured fields.

[0167] Example of a preset prompt template:

[0168] "You are the intelligent agent for reviewing machining process routes. Based on the given review unit, initial process route, and review evidence, please determine whether there are any problems in each route segment of the initial process route, and output the structured review results."

[0169] [Judging Unit]

[0170] Review type: {Review type}

[0171] Review target: {Review target}

[0172] Route segment awaiting review: {Route segment awaiting review}

[0173] Relevant constraint information: {Relevant constraint information}

[0174] [Initial Process Route]

[0175] {Initial Process Route}

[0176] [Evidence Review]

[0177] {Evidence Review}

[0178] Output Requirements

[0179] Please output the structured review results based on only the following fields:

[0180] Review type:

[0181] Review subjects:

[0182] Review conclusion:

[0183] Route problem type:

[0184] Problem location:

[0185] Reason for the problem:

[0186] Source of evidence:

[0187] Suggested solution:

[0188] Example of structured review results for each review unit:

[0189] ① The structured review results of the positioning hole machining integrity review unit in step 4.3: {

[0190] "Review Type": "Process Completeness"

[0191] "Review Target": "Location Holes"

[0192] Review Conclusion: "Problems Exist",

[0193] "Route Problem Type": "Insufficient Process Detailing"

[0194] Problem location: "Reaming positioning hole process",

[0195] "Cause of the problem": "The required size of the positioning hole is Φ8H7, which is a high-precision hole machining requirement; the initial process route has set up drilling and reaming of the positioning hole, but the finishing of the positioning hole only uses a single reaming process, without refining the reaming process into rough reaming and fine reaming."

[0196] Source of evidence: Article 14 of the Drilling Technique Code.

[0197] Recommended solution: Adjust the reaming process for the positioning holes to include both rough reaming and fine reaming.

[0198] },

[0199] ②Structured review results of the positioning hole quality control review unit in step 4.3:

[0200] {

[0201] "Review Type": "Quality Control"

[0202] "Review Target": "Positioning Hole Position Accuracy Inspection"

[0203] Review Conclusion: "Problems Exist",

[0204] "Route Problem Type": "Missing Inspection Procedure"

[0205] Problem Location: Inspection stage after final machining of positioning holes

[0206] "Cause of the problem": "The positional accuracy requirement of the positioning hole relative to the reference plane is 0.03mm, and the initial process route did not include a positioning hole positional accuracy detection step."

[0207] Source of evidence: "Application Case of 'Method for Detecting the Position of Thread Locating Holes using a Three-Coordinate Measuring Tool'",

[0208] "Suggested approach": "After the final machining of the positioning holes is completed, perform a three-coordinate measuring machine (CCM) check to verify the position of the positioning holes."

[0209] },

[0210] ③ The structured review results of the reference plane machining process sequence review unit in step 4.3:

[0211] {

[0212] "Review Type": "Process Sequence"

[0213] "Review Object": "The sequence of operations for rough milling the datum plane and finish milling the datum plane",

[0214] Review Conclusion: "No issues found".

[0215] "Route Problem Type": "None"

[0216] "Problem Location": "None"

[0217] "Cause of the problem": "The initial process route was set up in the order of rough milling reference plane first and finish milling reference plane second, which conforms to the process principle of leaving machining allowance for finish machining and ensuring that the design requirements are met after finish machining."

[0218] Source of evidence: Article 5.6 of the "General Technical Specifications for Cutting Processes".

[0219] Recommended handling: Do not adjust this route segment.

[0220] },

[0221] The structured review results of the hinge positioning hole resource matching review unit in step 4.3:

[0222] {

[0223] "Review Type": "Resource Matching"

[0224] "Review Target": "Locating Hole Machining Resources"

[0225] Review Conclusion: "No issues found".

[0226] "Route Problem Type": "None"

[0227] "Problem Location": "None"

[0228] "Cause of the problem": "The existing resources, including machining centers, special fixtures, and reamers, can cover the main equipment, fixtures, and tool types required for positioning hole clamping, pre-drilling, and reaming; in historical process cases, similar assembly positioning holes were clamped using machining centers with special fixtures and reamed using reamers."

[0229] Source of evidence: Fragments of historical craft cases

[0230] Recommended handling: Do not replace processing resources.

[0231] }

[0232] Step 5: Confirm or adjust the initial process route based on the structured review results, and generate the target process route.

[0233] In step 4.4 of this embodiment, the integrity review result for the positioning hole machining is problematic, the route problem type is insufficient process refinement, the problem location is the reaming positioning hole process, and the recommended solution is to adjust the reaming positioning hole to a coarse reaming positioning hole and a fine reaming positioning hole; the quality control review result for the positioning hole is problematic, the route problem type is missing inspection process, the problem location is the inspection stage after the final machining of the positioning hole, and the recommended solution is to add a three-coordinate measuring machine (CCM) inspection of the positioning hole position after the final machining of the positioning hole; the review results for the datum plane machining process sequence and the reaming positioning hole resource matching are both problem-free, and no adjustment to the corresponding route segments is required. Meanwhile, it is assumed that the review results for the four types of review units for the remaining machining objects are all problem-free.

[0234] Based on the structured review results above, the "reaming positioning hole" in the initial process route obtained in step 3.3 is replaced with "rough reaming positioning hole" and "fine reaming positioning hole". After the positioning hole is finally machined, the "three-coordinate measuring machine to check the positioning hole position accuracy" process is added, resulting in the target process route as follows: rough milling of the reference plane → fine milling of the reference plane → rough milling of the weight reduction groove → fine milling of the weight reduction groove → drilling of the positioning hole → rough reaming of the positioning hole → fine reaming of the positioning hole → three-coordinate measuring machine to check the positioning hole position accuracy → drilling of the bolt mounting hole.

[0235] Step 6: Provide feedback on the review results, categorize them, and supplement the knowledge.

[0236] Step 6.1: Read the causes of problems and sources of evidence from the structured review results, determine the knowledge update category according to the preset classification rules, and generate corresponding knowledge update candidates:

[0237] In this embodiment, the problem in the integrity review result of the positioning hole machining is "failure to distinguish between rough reaming and fine reaming when reaming high-precision holes". The suggested solution is "adjust the reaming positioning hole to both rough reaming and fine reaming positioning holes". Therefore, the problem cause and suggested solution are converted into rule base update candidates: applicable to high-precision positioning holes, triggered by machining features of positioning holes, high dimensional accuracy requirements, and the use of reaming as the finishing method for positioning holes, with the constraint that rough reaming and fine reaming positioning holes should be set after drilling the positioning hole, and the evidence source is Article 14 of the "Drilling Process Code".

[0238] The issue identified in the quality control review of the positioning holes was that "the positioning holes have positional accuracy requirements, but the initial process route did not include a corresponding positional accuracy detection step." The recommended solution was to "add a coordinate measuring machine (CMM) inspection step for the positioning holes after final machining." This issue reflects a lack of resource capability in the mechanical product manufacturing process knowledge graph regarding "the ability of CMM equipment to detect the positional accuracy of positioning holes." Therefore, the cause of this issue, the recommended solution, and the source of evidence are converted into process knowledge graph completion candidates: the completion entities include CMM equipment and positioning hole positional accuracy, and the hole positional accuracy detection step; the completion relationship is "hole positional accuracy detection step - used to detect - positioning hole positional accuracy, CMM equipment - able to detect - positioning hole positional accuracy," and the source of evidence is the application case of "Method for CMM Inspection of Threaded Positioning Hole Positional Accuracy."

[0239] The review results for the machining sequence of the reference plane and the resource matching of the hinge positioning hole were both found to be without problems, so no update candidates were generated.

[0240] In addition, in step 5, a process route adjustment record is generated based on the results of the processing integrity review and the quality control review, and the adjusted target process route is used as a candidate for process case update.

[0241] Step 6.2: The process engineer reviews the candidates obtained in Step 6.1 and provides a manual confirmation result. The manual confirmation result includes confirming the writing, writing after modification, and not writing.

[0242] Step 6.3: Based on the manual confirmation results, convert the confirmed rule base update candidates into rule items containing applicable objects, triggering conditions, constraint relationships, and evidence sources, and write them into the mechanical product manufacturing process rule base; convert the confirmed knowledge graph completion candidates into nodes, relationships, or attributes, and write them into the mechanical product manufacturing process knowledge graph; and write the confirmed process case update candidates into the retrieval enhancement review knowledge base after vectorization.

Claims

1. A method for generating a process route based on a knowledge graph and an evaluation agent, characterized in that, Including the following steps: Step 1: Construct a knowledge graph of mechanical product manufacturing processes, a rule base for mechanical product manufacturing processes, and a knowledge base for enhanced retrieval and review. The knowledge base for enhanced retrieval and review consists of original text fragments extracted from mechanical product process data that can serve as the basis for process route review, text vectors after the original text fragments are segmented and vectorized, metadata records corresponding to the original text fragments, vector indexes of the text vectors, and the association between the text vectors and metadata records. Metadata records include the document name, clause location, and version information. Step 2: Convert the processing information of the object to be processed into standardized process input data, including processing characteristics, quality requirements, resource constraints, and the relationship between quality requirements and processing characteristics or the object to be processed as a whole; Step 3: Generate the initial process route based on standardized process input data, mechanical product manufacturing process knowledge graph, and mechanical product manufacturing process rule base; Step 4: Construct a review agent based on retrieval enhancement generation technology and large language model, and use the review agent to review the initial process route and generate structured review results. The review agent comprises a review unit construction module, a retrieval query statement generation module, a review evidence generation module, a review prompt information generation module, and a large language model, used to implement the following steps: Step 4.1: The review unit construction module divides the initial process route into multiple review units for different review objects according to four review types: process integrity, quality control, process sequence, and resource matching. The review unit includes review type, review object, route segment to be reviewed, constraint information, and review purpose. Step 4.2: The query statement generation module generates a corresponding query statement for each review unit according to the preset query template. Step 4.3: For each review unit, the review evidence generation module uses its corresponding search query to retrieve review evidence from the enhanced review knowledge base, and sorts the retrieved review evidence according to its relevance to the current review unit, taking the top 10% to constitute review evidence. Step 4.3: The review prompt information generation module combines the initial process route, review unit, and review evidence into review prompt information for each review unit. Step 4.4: The large language model reviews each review unit separately: The review prompt information corresponding to the review unit is input into the large language model to determine the process route problem. The large language model outputs the structured review results of the current review unit, including the review conclusion, route problem type, problem location, problem cause, source of evidence and suggested handling method. Step 5: Generate the target process route based on the structured review results: When the review conclusion indicates that there are no problems with the initial process route, the initial process route is used as the target process route; when the review conclusion indicates that there are problems with the initial process route, the corresponding route segment in the initial process route is modified according to the location of the problem and the suggested handling method to obtain the target process route. Step 6: Complete knowledge based on structured review results: Classify the causes and sources of evidence in the structured review results, and convert reusable rule issues, graph knowledge issues and case issues into candidate options for updating the process rule base, candidate options for completing the process knowledge graph, and candidate options for updating the process case, respectively. After manual confirmation, write the candidate options into the corresponding mechanical product manufacturing process knowledge graph, mechanical product manufacturing process rule base or retrieval enhancement review knowledge base. 2.The knowledge graph and review agent-based process route generation method of claim 1, wherein, In step 1: The method for constructing a knowledge graph of mechanical product manufacturing processes is as follows: extract text from mechanical manufacturing process standards, historical process cases of mechanical products, and mechanical product process experience documents, and then perform segmentation and terminology standardization on the extracted text in sequence. Extract process object entities, process activity entities, and resource detection entities from the normalized text; identify the control relationships, resource matching relationships, and process sequence relationships between entities; Entities are converted into nodes and relationships into edges to form a knowledge graph of mechanical product manufacturing processes; The method for constructing the mechanical product manufacturing process rule base is as follows: extract process constraints from mechanical manufacturing process standards, historical process cases of mechanical products, and mechanical product process experience documents, and convert the process constraints into different rule items containing applicable objects, triggering conditions, and constraint relationships. These rule items constitute the mechanical product manufacturing process rule base. The method for constructing the retrieval-enhanced review knowledge base is as follows: Mechanical manufacturing process standards, audited manufacturing site issues and handling records, and historical process cases are used as review knowledge sources. Original text fragments that can serve as the basis for process route review are extracted from these sources, and metadata records are established for each fragment. The original text fragments are then segmented according to semantic integrity, and the resulting text is vectorized to obtain text vectors. These text vectors are associated with and stored in relation to their corresponding metadata records, and a vector index is established based on the text vectors. Finally, the original text fragments, text vectors, metadata records, vector indexes, and the association between text vectors and metadata records together constitute the retrieval-enhanced review knowledge base. The text vectors and vector indexes are stored in the Milvus vector database. 3.The knowledge graph and review agent-based process route generation method of claim 1, wherein, Step 2 is as follows: Step 2.1: Extract and standardize the processing information of the object to be processed, and then collect the standardized information into the feature description field, quality labeling field, and resource restriction field to obtain structured processing information. Step 2.2: Based on the pre-built process identification dictionary and identification rules, identify processing features, quality requirements and resource constraints from the structured processing information obtained in Step 2.1, and establish the correlation between quality requirements and processing features or the whole object to be processed; Step 2.3 involves encapsulating the identified processing features, quality requirements, resource constraints, and the relationship between quality requirements and processing features or the object to be processed as a whole into standardized process input data.

4. The method for generating processing routes based on knowledge graphs and review agents according to claim 3, characterized in that, The process identification dictionary in step 2.2 is used to extract processing feature information, quality requirement information, and manufacturing resource information from the processing information. The process identification dictionary is constructed based on the mechanical product process data and includes a processing feature dictionary, a quality requirement dictionary, and a manufacturing resource dictionary. The identification rules are used to associate the identified processing feature information with the quality requirement information. The identification rules are constructed based on the mechanical product process data.

5. The method for generating processing routes based on knowledge graphs and review agents according to any one of claims 1-4, characterized in that, Step 3 specifically involves: Step 3.1: Using the processing features and quality requirements in the standardized process input data obtained in Step 2 as search conditions, search for matching initial candidate processes in the mechanical product manufacturing process knowledge graph constructed in Step 1, and filter out candidate processes in combination with resource constraints. Step 3.2: Associate the selected candidate processes with their corresponding processing characteristics, quality requirements, and resource constraints to form a set of candidate processes; Step 3.3: Call the mechanical product manufacturing process rule library established in Step 1 to establish constraint relationships for each candidate process in the candidate process set, including sequential relationships, dependency relationships and prohibition relationships, so as to organize the scattered candidate processes into a process directed graph; Step 3.4: Perform topological sorting or constraint solving on the directed graph of the process to obtain the initial process route.

6. The method for generating processing routes based on knowledge graphs and review agents according to claim 5, characterized in that, The preset query templates in step 4.2 include a processing integrity query template, a quality control query template, a process sequence query template, and a resource matching query template, each corresponding to one of the four review units; the specific details of each query template are as follows: Process integrity query template: Material / part type + Processing features + Quality requirements + Current existing processes + Processing processes and process integrity requirements; Quality control query template: Processing characteristics + Quality requirements + Existing inspection / control procedures + Quality control and inspection procedure requirements; Work process sequence query template: Current route segment + Relationship between adjacent work processes + Work process sequence requirements; Resource matching query template: Process to be reviewed + Quality requirements + Matched resources + Resource capabilities meet requirements.

7. The method for generating processing routes based on knowledge graphs and review agents according to claim 6, characterized in that, The large language model in step 4.4 uses Qwen2.5-72B-Instruct.

8. The method for generating processing routes based on knowledge graphs and review agents according to claim 6, characterized in that, In step 6: the candidate options for updating the process rule base include the applicable objects, triggering conditions, constraint relationships, and evidence sources; the candidate options for completing the process knowledge graph include entities, relationships, and evidence sources; and the candidate options for updating the process case include process route adjustment records and the target process route generated after the process route adjustment.

Citation Information

Patent Citations

  • Complex part process route generation and distribution method based on dynamic knowledge graph

    CN119378350A

  • Aircraft process instruction generation-self-review-correction method guided by large language model

    CN120973923A

  • A knowledge-driven automatic generation method for a part machining process and related products

    CN121303313B