Anomaly Detection based on Representation Learning for normal and faulty classification

Abstract

We present a data-driven anomaly detection system based on representation learning for normal and faulty classification. The proposed system extracts bottle-neck features, which contribute to discriminating the healthy and faulty states of machinery, through a deep neural network, and then detects an anomaly by a verifier constructed on the bottle-neck feature space. Experimental comparisons conducted using vibration data observed at wind turbines demonstrated that the bottle-neck features yielded improvements in anomaly detection over the conventional feature.