基于过程变量因果关系的炼化分馏装置异常检测方法

Anomaly Detection Method for Refining and Fractionation Equipment Based on Causal Relationship of Process Variables

以炼化分馏装置为对象,针对其过程变量间关联关系复杂且未知的特点,提出了一种基于过程变量因果关系的异常检测方法。该方法首先对装置历史正常运行数据进行格兰杰因果分析,构建反映过程变量关联关系的格兰杰因果系数矩阵,在此基础上训练图注意力网络,学习分馏装置的正常运行模式,最后结合阈值分析过程变量偏离正常运行状态的程度实现异常检测。上述方法在催化裂化分馏仿真系统上验证了有效性,并与三种常用的异常检测算法进行对比,结果表明该方法具备高效的异常检测能力,能为分馏装置系统运维人员提供参考。

This paper focuses on the catalytic cracking fractionation unit,and proposes an anomaly detection method based on causal relationship mining of process variables to address the complex and unknown correlation between process variables.This method first performs Granger causality analysis on the historical normal operation data of the device to construct a Granger causality coefficient matrix that reflects the correlation between process variables.Based on this,a graph attention network is trained to learn the normal operation mode of the fractionation device.Finally,combining threshold analysis to determine the degree of deviation of process variables from normal operating mode,anomaly detection is achieved.The above method has been validated for effectiveness on a catalytic cracking fractionation simulation system and compared with three commonly used anomaly detection algorithms.