Off-Line Detection of Abrupt and Transitional Change Points in Industrial Process Signals

Abstract

Off-line change point detection (CPD) methods are frequently used as a primary step for the retrospective diagnosis of abnormal situations. However, detecting both abrupt and more ambiguous transitional change points (CPs) in industrial process data remains a challenging task. To address and solve this limitation, we present a novel off-line CPD method that uses the "nonparametric multiplicative regression-based causality estimator" to calculate the bidirectional prediction residuals of a process signal. These residuals are then used as input to a "median absolute deviation"-based outlier detection and the "pruned exact linear time" CPD algorithm. Eventually, the resulting CPs are aggregated. Our method is compared to eight state-of-the-art methods. The detection performance of the evaluated methods is assessed using an openly accessible dataset based on the "Tennessee-Eastman-Process". It is demonstrated that our proposed method is beneficial for the off-line CPD performance of abrupt and transitional CPs.