CN-121980925-A - Forging process planning large model construction method and system based on industry specific corpus fine adjustment

CN121980925ACN 121980925 ACN121980925 ACN 121980925ACN-121980925-A

Abstract

The invention belongs to the technical field of industrial artificial intelligence and intelligent manufacturing, and particularly relates to a forging process planning large model construction method and system based on industry specific knowledge corpus fine tuning. The method comprises the following steps of firstly constructing a forging industry specific corpus, collecting historical process data, equipment operation parameters, material performance data and expert experience knowledge of a forging production line of a specific enterprise, cleaning and standardizing, secondly selecting a basic large model adapting to an industrial scene, designing a targeted fine tuning strategy, carrying out data enhancement by adopting migration learning and a generating type countermeasure network, carrying out parameter optimization based on a QLoRA efficient fine tuning framework, constructing a multi-objective optimization loss function, and finally deploying the model to a PLC control system or a WINCC monitoring platform of the forging production line, and realizing the intellectualization and refinement of process decision by OPC UA protocol real-time interaction data. The invention obviously improves the accuracy and efficiency of process decision making, and is suitable for various forging scenes.

Inventors

SU ZHENHUA
HE BO
WU LIANG
FAN YULIN
MA HAIKUAN
Ye Muqin
MA YONGJUN
FENG DONGXIAO
FU YONGTAO

Assignees

中国重型机械研究院股份公司

Dates

Publication Date: 20260505
Application Date: 20251231

Claims (7)

1. A forging process planning large model construction method based on industry specific corpus fine tuning is characterized by comprising the following steps: Step S1, a forging industry specific corpus is built, namely, historical process data, equipment operation parameters, material performance data and expert experience knowledge of a forging production line of a specific enterprise are collected, unstructured expert experience knowledge is converted into structural rules by utilizing a natural language processing technology through data cleaning and standardization processing, and association labeling is carried out to form the forging industry specific corpus containing the enterprise specific knowledge; S2, selecting a basic large model adapting to an industrial scene, namely selecting a general knowledge vertical domain large model with numerical process parameter processing capability and industrial text understanding capability as the basic large model; S3, designing a targeted fine tuning strategy of the forging process, namely expanding a sample by adopting a data enhancement technology aiming at scarce process data, constructing a multi-objective optimization loss function taking forging precision, production efficiency and cost control as optimization targets based on a parameter efficient fine tuning framework, and performing targeted fine tuning on a basic large model; s4, model training and verification, namely dividing a specific recognition corpus into a training set, a verification set and a test set, optimizing model parameters through iterative training, verifying process decision accuracy and production line fitness, and forming a specific recognition vertical domain large model; And S5, model deployment and iterative optimization, namely deploying the trained special knowledge vertical domain large model to a control system or a monitoring platform of the forging production line, receiving production line operation data in real time, and continuously updating model parameters in an online incremental learning mode.
2. The forging process planning large model construction method based on industry specific corpus fine tuning is characterized in that in the step S1, historical process data, equipment operation parameters, material performance data and expert experience knowledge of a specific enterprise forging production line are collected, and concretely comprises forging raw material chemical components and mechanical property data, forging process parameters, operation parameters of a forging manipulator and heating equipment, heat treatment process parameters, forging quality detection data and heavy machinery forging industry standard specifications, wherein the association label is a label for establishing a mapping relation among material characteristics, equipment parameters, process parameters and forging quality.
3. The forging process planning large model construction method based on industry specific corpus fine tuning of claim 1, wherein in the step S3, the data enhancement technology comprises the steps of migrating similar process characteristics by using a migration learning method aiming at scarce process data of a large complex forging, generating complementary data conforming to process distribution by combining a generated countermeasure network, and fine tuning the model by adopting a quantized low-rank adaptation technology through freezing basic model parameters and training a low-rank adaptation matrix by the parameter efficient fine tuning framework.
4. The forging process planning large model construction method based on industry specific corpus fine tuning of claim 1, wherein in the step S3, the multi-objective optimization loss function is represented by L=alpha.L_quality+beta.L_efficiency+gamma.L_cost, wherein L_quality is mass loss related to forging dimensional accuracy errors and internal defects, L_efficiency is efficiency loss related to production periods, L_cost is cost loss related to material utilization and energy consumption, and alpha, beta and gamma are weight coefficients dynamically adjusted according to production priorities.
5. The forging process planning large model construction method based on industry specific corpus fine tuning of claim 1 is characterized in that in the step S5, a trained specific vertical domain large model is deployed to a control system or a monitoring platform of a forging production line, an edge computing architecture is deployed, the model is deployed to an edge computing node of a production line PLC control system or a WINCC monitoring platform, data interaction is conducted between the model and production line equipment and between the model and a digital twin model through OPC UA protocol, and the online incremental learning mode comprises periodically collecting new process data and quality feedback of the production line, updating the specific corpus and conducting incremental fine tuning.
6. An intelligent decision-making system for forging process, which is a large model of forging process planning proprietary domain constructed by using the method for constructing large model of forging process planning based on fine adjustment of industry proprietary corpus as set forth in claim 1, comprising: The data acquisition module is used for acquiring real-time operation data and raw material information of the forging production line; The special-recognition vertical domain large model module internally runs a forging process planning special-recognition vertical domain large model and is used for generating a process planning scheme and a parameter optimization instruction according to the acquired data; The execution control module is connected with the production line PLC system and is used for receiving and executing the parameter optimization instruction; the quality feedback module is used for collecting the forged quality data and feeding the forged quality data back to the special-recognition vertical domain large model module for iterative optimization.
7. The intelligent decision system of forging process according to claim 6, wherein the intelligent decision system of forging process is configured for process planning of 300MN and above tonnage forging manipulator, and the specialized vertical domain large model module is configured with differential process route decision logic for shaft, frame and connecting rod forgings.

Description

Forging process planning large model construction method and system based on industry specific corpus fine adjustment Technical Field The invention belongs to industrial artificial intelligence and intelligent manufacturing technology, and particularly relates to a forging process planning large model construction method and system based on industry specific knowledge corpus fine tuning. Background With the rapid development of industry 4.0 and intelligent manufacturing, artificial intelligence technology is increasingly applied in the field of heavy machinery forging. The forging process planning is a key link for determining the quality (such as dimensional accuracy and mechanical property) and production efficiency of the forging, and the traditional method mainly relies on manual experience to compile process cards or performs verification based on Finite Element Modeling (FEM), so that the problems of low efficiency, high trial-and-error cost and difficulty in inheriting expert implicit knowledge exist. In recent years, data-driven based machine learning and digital twinning techniques have begun to be applied to forging process control. For example, chinese patent CN120145704B discloses a thermal processing control method based on digital twinning, which collects full-flow data through multi-sensor fusion, constructs a digital twinning model and deploys the model to an edge computing gateway, and calculates target process parameters by using a distributed optimization engine. However, such methods mainly rely on physical simulation models or numerical optimization algorithms, lack effective understanding and fusion capabilities for unstructured expert experiences (such as process recipe, fault handling notes), have complex modeling processes, and are difficult to quickly adapt to different production lines of different enterprises. On the other hand, the Large Language Model (LLM) is used at the beginning of the application in the fields of material science and industry. Shaohan Tian et al in 2024, paper STEEL DESIGN Based on a Large Language Model, propose a steel design method based on a large language model, which uses SteelBERT model to build an end-to-end prediction pipeline from material text description to mechanical properties, and realizes the composition design of specific steel types (such as 15 Cr) through fine tuning strategies. However, this prior art focuses mainly on the static mapping of "material composition and properties" and fails to address the complex process planning problems that involve equipment operating parameters (e.g., roll-down, operator load), dynamic thermal processes, and multi-objective constraints (cost, efficiency, quality) in the forging production process. Furthermore, general large model tuning methods typically target text generation consistency, lack constraint mechanisms for industrial numerical parameter accuracy, resulting in model generated process parameters that tend to be logically smooth but not engineering viable. In summary, the prior art has the following defects that firstly, a knowledge model or a material design model lacks adaptation to specific equipment parameters and process habits of a specific forging enterprise, secondly, an effective enhancement means for processing a scarce process sample (such as a large complex forging) is lacking, thirdly, the existing model training does not directly incorporate production cost, efficiency and quality indexes into a loss function for multi-objective optimization, and is difficult to directly guide the fine production of an actual production line. Therefore, there is a need for a large model construction method for forging process planning vertical domain that can integrate industry specific identification, fit a specific production line, and have multi-objective decision making capability. Disclosure of Invention Aiming at the problems that a general large model lacks specific process special knowledge of a forging enterprise, is difficult to adapt to the fine decision requirement of different production lines, a traditional model is poor in performance under a scarce process data scene, and the calculation cost of a general fine tuning method is high and is not optimized for industrial multiple targets in the prior art, the invention aims to provide a forging process planning large model construction method and a system based on industry special knowledge corpus fine tuning. In order to achieve the above purpose, the invention provides a forging process planning large model construction method based on industry specific knowledge corpus fine tuning, which comprises the following steps: Step S1, a forging industry specific corpus is built, namely, historical process data, equipment operation parameters, material performance data and expert experience knowledge of a forging production line of a specific enterprise are collected, unstructured expert experience knowledge is converted into structural ru