Binary Classifier for Fault Detection Based on KDE and PCA

Abstract

Fault detection in process condition monitoring aims to declare anomalies strayed from the operation expectancy. With the number of variables increasing, the complexity of detection task grows quickly. The algorithm, Binary Classifier for Fault Detection (BaFFle), is devised to employ Principal Component Analysis (PCA) to reduce the number of variables and to detect the occurrences of fault over each distinct component using corresponding operation model. In this way, BaFFle converts a multivariate detection task into several univariate problems. Since the observations of a steady-state system are supposed to subject to certain probability distribution, the normal operation model is represented by probability distribution. In the original BaFFle algorithm, measured data is assumed Gaussian distribtued. In order to get rid of the strong assumption in BaFFle, an improved BaFFle is proposed in this paper to estimate the distribution model by Kernel Density Estimation (KDE). A main advance of KDE is attributed to the non-parametric way of estimating probability distribution, resulting in a data-driven estimator. Experiments using real data from a multiphase flow rig proved the validation of KDE-based BaFFle. From the practice perspective, BaFFle has the potential of being applied to operation systems with non-Gaussian distributions.