Toward Self-Supervised Feature Learning for Online Diagnosis of Multiple Faults in Electric Powertrains

Abstract

This article proposes a novel online fault diagnosis scheme for industrial powertrains without using historical faulty or labeled training data. The proposed method combines a one-class support vector machine (SVM) based anomaly detection and supervised convolutional neural network (CNN) algorithms to online detect multiple faults and fault severities under variable speeds and loads. The one-class SVM algorithm is to derive a score for defining faults or health classes in the first stage, and the resulting health classes are used as the training data for the CNN-based classifier in the second stage. Within this framework, the self-supervised learning of the proposed CNN algorithm allows the online diagnosis scheme to learn features based on the latest data. The effectiveness of the scheme is validated via a comparison study using experimental data from an in-house test setup. Finally, the online implementation of the proposed scheme on the test setup is briefly introduced.