基于正交局部慢性特征的故障检测方法

Fault detection based on orthogonal local slow features

为提高化工行业中数据驱动故障检测的效果,该文针对实际工业系统中闭环控制导致的过程动态特性和数据流形中蕴含的局部信息,提出了一种基于局部时空正则慢特征提取(local time-space regularized slow feature extraction, LTSS)的方法进行故障检测。首先,构造基于局部时空正则的目标函数得到投影矩阵,进而得到预提取特征S,则S张成的空间中包含了静态信息,而S的一阶差分张成的空间中包含了动态信息。其次,基于独立成分分析(independent components analysis, ICA)方法,分别为2个空间构建对应的S^2和SPE统计量进行监控,用于实时故障检测。在TE (Tennessee Eastman)过程上的案例研究可以证明所提方法的有效性。

A local time-space regularized slow feature extraction method was developed to improve data-driven fault detection in the chemical industry based on the process dynamics of closed-loop control systems and the local information contained in the data manifold.An objective function was defined based on the local time-space term to obtain a projection matrix and the pre-extraction feature,S.The span of S contains the static information,while the first derivative of the span of S contains the dynamic information.An independent component analysis was used to obtain statistics for S2 and SPE for both spaces for real-time fault detection.A case study on the Tennessee Eastman process shows the validity of this method.