Abstract:

For fault detection in dynamic processes, a novel dimensionality reduction method was proposed on the basis of improved neighbor selection in neighborhood preserving embedding algorithm, i.e., time constrained neighborhood preserving embedding(TCNPE). Compared to neighborhood preserving embedding(NPE), which selected neighborhood only by Euclidean distance, TCNPE selected neighborhoods within certain timeframe by k-nearest neighboring method with a consideration of time series correlation between data points and constructed a localized constraining relationship between near and distant neighborhoods within this time window. First, TCNPE algorithm extracted main features of the process data and performed linear dimensionality reduction. Next, Hotelling's T² and SPE statistics were established for online process monitoring and kernel density estimation(KDE) was used to determine control limits. Case study and simulation of Tennessee-Eastman Process demonstrated efficacy of the proposed method.

Key words: process control, dynamic modeling, neighborhood preserving embedding, linear dimensionality reduction, fault detection, experimental validation

摘要：

针对动态过程的故障检测问题，在邻域保持嵌入算法中改进邻域挑选，提出一种新的维度约简方法：时序约束邻域保持嵌入（time constrained neighborhood preserving embedding，TCNPE）算法。与邻域保持嵌入（neighborhood preserving embedding，NPE）算法只通过欧氏距离挑选邻域不同的是，TCNPE考虑到数据之间的时序相关性，在一定长度的时间窗之内采用k-近邻方法挑选邻域，并对时间窗内近邻与非近邻构造局部约束关系。首先，利用TCNPE提取数据特征，进行线性降维，然后构造T²和SPE统计量并利用密度估计（kernel density estimation，KDE）确定其控制限。最后，通过数值例子和TE过程（Tennessee-Eastman process）仿真来说明本文方法的有效性。

关键词: 过程控制, 动态建模, 邻域保持嵌入, 线性降维, 故障检测, 实验验证