| 摘要: |
| 目的 含硫天然气净化生产易燃易爆、连续且过程复杂，安全风险大，故障归因与溯源对操作人员排查隐患、预防事故和保障安全生产至关重要，对工程人员操作有重要指导意义。知识图谱可高效存储管理化工生产资料，为故障溯源等任务提供数据支持，提升运维效率。但现有生产运维资料多为非结构化文本，限制了知识图谱的构建。针对此问题，提出了一种双向长短期记忆网络（BiLSTM）与条件随机场（CRF）融合规则匹配的知识抽取方法。方法 首先采集工业过程的生产资料或运维资料，作为原始数据并进行预处理，接下来利用BiLSTM-CRF和规则匹配相结合的方法进行知识抽取，将抽取的数据存储于图数据库中。结果 以天然气净化厂闪蒸罐为例，使用该方法构建的知识图谱与专家经验构建的理论图谱结构基本一致。结论 实验结果表明，所提出的模型能有效地提取装置的生产资料或运维资料中的知识。构建的知识图谱增强了资料的可读性，便于运维人员查询和学习。 |
| 关键词: 天然气净化 长短时记忆网络 条件随机场 命名实体识别 知识抽取 知识图谱 |
| DOI：10.3969/j.issn.1007-3426.2025.03.020 |
| 分类号: |
| 基金项目:国家自然科学基金项目“基于图网络的复杂过程系统深度强化学习方法研究”（62273026） |
|
| Rule-based knowledge extraction method for typical equipmentin natural gas purification |
|
JI Tianhao1, PENG Chuanbo2, PEI Aixia2, ZHOU Jian2, LIU Chiqiang1, LI Dazi1
|
|
1.College of Information Science and Technology, Beijing University of Chemical Technology, Beijing, China;2.Zhongyuan Oilfield Puguang Branch, Dazhou, Sichuan, China
|
| Abstract: |
| Objective The production process of natural gas purification is characterized by flammability, explosiveness, continuity, and complexity, all of which pose substantial safety risks. Fault attribution and tracing are crucial for operators to identify potential hazards, prevent accidents, and guarantee safe production. Furthermore, they provide invaluable guidance for engineering personnel in their operations. Knowledge graphs can efficiently store and manage the vast amount of process technical regulations and fault handling logs in chemical production, offering data support for downstream tasks such as fault tracing and enhancing the efficiency of maintenance personnel. However, most existing production and maintenance data are recorded and stored in unstructured text data, curtailing the potential for directly constructing a knowledge graph. To address this challenge, this study proposes a knowledge extraction method that combines a bidirectional long short-term memory network (BiLSTM) with conditional random field (CRF) and rule matching. Method Firstly, production or operation and maintenance data pertaining to relevant industrial processes are gathered as original data and subjected to preprocessing. Secondly, the method integrating BiLSTM-CRF and rule matching was employed for knowledge extraction. Finally, the extracted data was stored in the graph database. Result In this study, the flash tank device in a natural gas purification plant was taken as an example. The constructed knowledge graph was basically consistent with the theoretical graph structure constructed by expert experience. Conclusion The proposed model can effectively extract knowledge about the device's production operation and maintenance data. The constructed knowledge graph enhances the data readability and facilitates ease of query and learning for operation and maintenance personnel. |
| Key words: natural gas purification bidirectional long short-term memory network conditional random field named entity recognition knowledge extraction knowledge graph |
