A novel information processing method based on an ensemble of Auto-Encoders for unsupervised fault detection

Abstract

The current research paper proposes a novel information processing Deep Learning framework for unsupervised fault detection applications, where during the training process only samples from the normal class are available. Recently, unsupervised fault detection methods based on the Auto-Encoders have been successful and are used extensively. They are supported by the assumption that abnormal unknown samples that don’t belong to the learned manifold of the training dataset of normal points produce higher reconstruction cost than the normal samples. The presented scheme is based on an ensemble of different types of Auto-Encoders; each of them is trained independently for the one-class fault detection task. The final agreement of the normality of the testing sample is made from a soft voting process, where the confidence of each Auto-Encoder about its decision is considered. The significance of each individual Auto-Encoder in the final decision is extracted from a statistical analysis of the independent training process of each one. Simulation results with three widely used fault detection datasets show the effectiveness of the proposed model.