Continual learning for fault diagnosis considering variable working conditions

Abstract

Traditional Neural Networks (NNs) trained in a one-stage process often struggle to perform well when presented with new classes or domain shifts in testing datasets. In fault diagnosis, it is essential to handle a sequence of diagnostic tasks with new fault classes and working conditions. This paper presents a multi-staged Continual Learning algorithm that learns from a sequence of diagnostic tasks. In each training stage, a small portion of previously seen training data is incorporated to help the model remember old tasks and better learn new tasks. A novel scheme is designed to select previously seen data from multiple old tasks, considering their different working conditions. A multi-way domain adaptation is then conducted to mitigate the impact of multiple changes in working conditions among different tasks. The proposed method is tested using two different experiment test rigs, including both gear and bearing faults. Results demonstrate that the proposed Continual Learning algorithm allows NNs to learn from a sequence of diagnostics tasks efficiently and maintain high accuracies for all the tasks of interest.