当前位置：首页 > 过刊浏览->2026年第52卷第4期
基于AVEVA和大模型推理的阀门物料表单自动生成技术
Automatic valve material list generation technology based on AVEVA and LLM reasoning
浏览(9) 下载(0)
- DOI:
- 作者:
- 陈 磊，雷俊杰，张斯祺，钟 旭，王亚月
Chen Lei, Lei Junjie, Zhang Siqi, Zhong Xu, Wang Yayue
- 作者单位:
- 海洋石油工程股份有限公司设计院，天津 300451
- 关键词:
- 大模型；智能P&ID；阀门物料表单；工程设计自动化；知识库
- 摘要:
- 在海洋石油平台管道设计阶段，阀门物料表单的生成通常依赖人工在P&ID中逐个识别阀门，反复查
阅《管道材料等级规格书》以确定阀门代码等隐含属性，工作效率低、差错率高，且难以适应设计文件频繁修
改的工程需求。针对阀门统计问题，提出一种面向工程设计流程的阀门物料表单自动生成方法，基于AVEVA
平台开发 PML 脚本并构建阀门、管线与仪表属性表，实现阀门实体及关键属性的结构化抽取，将规格文档构
建为矢量知识库与半结构化知识库，结合检索增强与约束校验机制，引入大模型作为统一推理引擎，设计AI
Chain推理链，实现阀门代码的自动推断与表单字段补全，最终输出满足工程规范的阀门物料表单。本文依托
包含多类典型复杂工况的工程数据完成模型测试验证，结果表明：场景级测试中120台阀门的整表均一致性校
验通过；在620组样本上阀门代码准确率达到98.55%，显著优于其他3种对比基线方法；误差分析中发现的
错误主要由项目定制条款未覆盖导致，后续可通过引入项目专用知识库与规则扩展进一步提升跨项目泛化能力
与生成可控性。自动生成技术可提高阀门物料表单统计结果的准确性、一致性和可审计性，从而提升设计工作
整体效率。
In the pipeline design stage of offshore oil platforms, the generation of valve material lists usually relies on engineers manually identifying valves one by one in the Piping and Instrumentation Diagram (P&ID) and repeatedly consulting the Piping Material Class Specification to determine implicit attributes such as valve codes, which results in low efficiency and high error rates and makes it difficult to adapt to frequent revisions of design documents. To address the valve statistics problems, this paper proposes an automatic valve material list generation method for engineering design workflows. Based on the AVEVA platform, PML scripts are developed, and valve, pipeline, and instrument attribute tables are constructed, thereby enabling the structured extraction of valve entities and key attributes. The specification documents are organized into a vector knowledge base and a semi-structured knowledge base. By combining retrieval augmentation and constraint checking mechanisms, a large language model (LLM) is introduced as a unified reasoning engine, and an AI Chain reasoning chain is designed to automatically infer valve codes and complete list fields. Finally, valve material lists that satisfy engineering specifications are output. The proposed method is tested on engineering data containing multiple types of complex working conditions. The results show that in the scenario-level test, the complete material lists of 120 valves pass the consistency check. For 620 samples, the valve code prediction accuracy reaches 98.55%, significantly outperforming the other three baseline methods. Error analysis shows that the remaining errors are mainly caused by the lack of coverage of project-specific clauses. In future applications, cross-project generalization and generation controllability can be further improved by introducing project-specific knowledge bases and rule extensions. The proposed automatic generation technology can improve the accuracy, consistency, and auditability of valve material list statistics, thereby enhancing the overall design efficiency.