欢迎访问过程工程学报, 今天是

过程工程学报 ›› 2022, Vol. 22 ›› Issue (1): 135-144.DOI: 10.12034/j.issn.1009-606X.221071

• 研究论文 • 上一篇    

基于极深因子分解机的化工过程故障诊断方法

何亚东,袁壮,林扬,高新江,李传坤*,王春利   

  1. 中国石化安全工程研究院化学品安全控制国家重点实验室, 山东 青岛 266071
  • 收稿日期:2021-03-01 修回日期:2021-05-08 出版日期:2022-01-28 发布日期:2022-01-28
  • 通讯作者: 李传坤 lick.qday@sinopec.com
  • 作者简介:何亚东(1993-),男,山东省潍坊市人,硕士研究生,工程师,研究方向为数据挖掘与分析、故障诊断;李传坤,通讯联系人,E-mail: lick.qday@sinopec.com.
  • 基金资助:
    基于深度学习的复杂化工过程异常模式智能识别研究

Chemical process fault diagnosis method based on extreme deep factorization machine

Yadong HE,  Zhuang YUAN,  Yang LIN,  Xinjiang GAO,  Chuankun LI*,  Chunli WANG   

  1. State Key Laboratory of Safety and Control for Chemicals, SINOPEC Research Institute of Safety Engineering, Qingdao, Shandong 266071, China
  • Received:2021-03-01 Revised:2021-05-08 Online:2022-01-28 Published:2022-01-28

摘要: 近年来,随着化学工艺、设备的复杂化和大型化程度不断深入,如何为化工企业及时、准确地诊断故障、排除事故,成为一个极具挑战性的问题。目前,以深度学习为代表的化工过程故障检测与诊断技术成为业界解决问题的主要思路之一,但现有深度方法在构建诊断模型时只关注了变量的非线性高阶特征,不能充分、全面地挖掘多源交互信息,难以有效地融合各类异构数据。基于此,提出一种基于极深因子分解机的化工过程故障诊断方法,通过并行融合三类不同网络模型(分解机模型、深度神经网络模型、压缩交互网络模型),实现对高阶、低阶及线性特征的自动提取和高效整合。为了评估模型性能,从单故障诊断和多故障混合诊断的角度出发,在田纳西-伊斯曼过程(TE)仿真数据集上进行了广泛的对比实验,结果表明,所提方法较以往故障诊断方法在精确率和召回率等指标上具有明显优势。

关键词: 故障检测与诊断, 深度学习, 极深因子分解机, 特征交互, 田纳西-伊斯曼过程

Abstract: The chemical process fault detection and diagnosis technology represented by deep learning has become one of the main ideas to solve the problem in the industry. However, the existing deep learning diagnosis methods only focus on the non-linear high-order interactive features when constructing training models and ignore the complementary of linear features and low-order interactive features to global modeling. In addition, the high-order features extracted by the existing deep models involve only implicit interactive features, whose feature forms are unknown and uncontrollable in order. Based on these problems, this work proposes a extreme deep factorization machine-based fault diagnosis method for chemical processes, which achieves automatic extraction and efficient integration of high-order, low-order and linear features by parallel fusion of three different types of network models (factorization machine, deep neural networks and compressed interaction network). First, the selected data are sequentially subjected to preprocessing operations such as Z-score normalization, label annotation, and format conversion to convert the input data into the format data required by the model. Then, the format data are simultaneously input to the three neural network models to help train the proposed diagnostic model in parallel. Finally, the fault diagnosis results are output based on the optimal diagnosis model. From the perspective of single-fault diagnosis and multi-fault hybrid diagnosis, extensive comparison experiments are conducted on the Tennessee-Eastman process (TE) simulation dataset, and the results show that the proposed method has significant advantages over previous fault diagnosis methods in terms of metrics such as precision and recall rate.

Key words: Fault detection and diagnosis, Deep learning, Extreme deep factorization machine, Feature interaction, Tennessee Eastman process