欢迎访问过程工程学报, 今天是

过程工程学报 ›› 2024, Vol. 24 ›› Issue (8): 937-945.DOI: 10.12034/j.issn.1009-606X.223280CSTR: 32067.14.jproeng.223280

• 研究论文 • 上一篇    下一篇

化工过程预警的CNN-LSTM耦合模型研究

崔劲松1,2， 贾波2， 李学盛2， 王亚茹2， 李海航1， 王海宁1*， 包其富2*   

  1. 1. 中国计量大学质量与安全工程学院，浙江 杭州 310018 2. 浙江省应急管理科学研究院，浙江 杭州 310061
  • 收稿日期:2023-10-18 修回日期:2024-01-15 出版日期:2024-08-28 发布日期:2024-08-22
  • 通讯作者: 王海宁 whnfyy@163.com
  • 基金资助:
    精细化工过程安全风险管控关键技术研究与应用;国家自然科学基金青年项目

Research on CNN-LSTM coupling model for chemical process early warning

Jingsong CUI1,2,  Bo JIA2,  Xuesheng LI2,  Yaru WANG2,  Haihang LI1,  Haining WANG1*,  Qifu BAO2*   

  1. 1. College of Quality and Safety Engineering, China Jiliang University, Hangzhou, Zhejiang 310018, China 2. Zhejiang Academy of Emergency Science and Technology, Hangzhou, Zhejiang 310061, China
  • Received:2023-10-18 Revised:2024-01-15 Online:2024-08-28 Published:2024-08-22
  • Supported by:
    Study and application of safety risk management and control technology in fine chemical process

RichHTML

PDF

摘要： 针对化工过程参数的实时性、多维性、非线性，以及化工流程复杂、互相干扰项多、预警方法单一等情况，本工作提出了一种针对化工过程异常工况的深度学习回归预测与ADF (Augmented Dickey-Fuller)检验相耦合的预警模型。在对缩合反应的超温异常工况监测预警分析中，使用卷积神经网络和长短期记忆网络(Convolution Neural Network-Long Short-Term Memory, CNN-LSTM)模型实现了对未来400 s内过程重要参数的预测。同时利用ADF检验，对该时段的重要参数进行趋势检验。当结果为趋势不稳定，并且CNN-LSTM模型预测到某时刻重要参数超过警戒线，再对安全人员发出警报。结果表明，本工作所提出的方法在检测到过程参数不平稳后，分别以模型拟合优度(R^2)为0.9827和0.9882的情况下提前18和16 s预测到异常工况的发生，从而实现化工过程异常工况的提前预警，对实现化工过程安全运行进行了有利探索。

关键词: 安全工程, 化工过程, 监测预警, 深度学习, ADF检验

Abstract: In consideration of the real-time, multi-dimensional, and nonlinear nature of chemical process parameters, as well as the complexity of chemical processes with numerous mutually interfering factors and single warning method, this work proposes an early warning method combining deep learning regression prediction and ADF (Augmented Dickey-Fuller) test. For monitoring and early warning analysis of over-temperature abnormal conditions in condensation reactions, convolutional neural network, and long short-term memory (CNN-LSTM) models are employed in this study to predict crucial process parameters for the next 400 s. Simultaneously, the ADF test is utilized to examine the trend of temperature time series parameters. When the result is an unstable trend and the CNN-LSTM model predicts that the temperature will exceed the alarm threshold at a specific time point, security personnel will be alerted accordingly. The results showed that during the condensation reaction's over-temperature anomalies at feed rates of 700 and 800 kg/h, the CNN-LSTM model's regression forecasting for temperature metrics manifested R2 values of 0.9827 and 0.9882. Correspondingly, the model elicits RMSE (Root Mean Square Error) values of 0.1425 and 0.1453, and MAE (Mean Absolute Error) values of 0.1184 and 0.1234. These indices testify to the model's exceptional fidelity and precision, surpassing the conventional LSTM model's predictive accuracy as reflected in its R2, RMSE, and MAE values. The ADF test results on the temperature time series data corroborate the presence of an unstable trend, aligning with the actual process behavior. By combining both methods, the early warning model is able to detect temperatures exceeding the alarm threshold 18 and 16 s earlier than the simulated alarm point, respectively, and issues a timely alert. The dual application of these methods provides a robust means of monitoring chemical process parameters, enabling the early detection of abnormal conditions in chemical processes and advancing the field of chemical process parameter monitoring.

Key words: safety engineering, chemical processes, monitoring and warning, deep learning, ADF test