Unknown fault diagnosis based on meta-learning under complex working conditions

Abstract

Traditional fault diagnosis methods are often based on multiple hidden layer networks and self-adaptive extract features from massive data, according to the data under specific working conditions, a specific model is learned. When faced with a small amount of unknown fault type data or complex working conditions, effective diagnosis cannot be performed. This paper proposes a meta-learning algorithm that uses known fault type data to build a neural network model during the training phase and determines a threshold, make a preliminary distinction between known faults and unknown faults through basic decision-making, and learn another more robust decision based on the effect of the distinction, and make the machine have the ability to learn to learn, and use the learned decisions under another working condition in the testing phase to also effectively diagnose unknown faults, thus avoiding retraining or learning in new tasks. The feasibility and accuracy of this method are verified in Western Reserve University Bearing Fault Database.