A multivariate Gaussian mixture model for anomaly detection in transient current signature of control element drive mechanism

Abstract

The analysis of transient coil current signals for anomaly detection in a control element drive mechanism (CEDM) involves identifying signature mismatch based on comparison to normal signal profiles. This paper presents an anomaly detection model for optimizing the health monitoring framework for CEDMs. Our approach is a machine learning clustering algorithm based on the multivariate Gaussian Mixture Model (GMM), for the detection of incipient abnormal transient signatures in CEDM coils. The underlying considerations for the choice of a GMM-based clustering approach are that (1) the transient coil current profiles at normal operational behavior are very similar, and (2) only very few data are usually available that describe anomalous CEDM behavior. Since available data is heavily skewed toward normal patterns, this methodology allows us to establish the norm for detecting anomalous patterns based on a preset detection threshold. The model evaluation result shows a very low false-negative rate for anomaly detection.