DRL-GCNet: A Deep Reinforcement Learning and Graph Convolutional Network for Harmonic Drive Fault Diagnosis

Abstract

Harmonic drives (HDs) are key components of industrial robots, and their malfunction or breakdown can cause robot operational mistakes. Therefore, accurately diagnosing faults of HDs is of great significance for their applications. In this article, fault diagnosis features of HDs were extracted from vibration signals using short-time Fourier transform (STFT) by creating spatiotemporal graphs, and a deep reinforcement learning (DRL) framework was employed to enhance diagnostic process, wherein agents were trained to assign structure of graph neural networks and aggregation strategies. Meanwhile, the ChebGCN was used to classify the fault characteristics of HDs under various fault states and working conditions. This was the first time that DRL and ChebGCN were jointly used for HD fault diagnosis. Comprehensive experiments were run to validate the effectiveness of the proposed methods on two separate datasets, wherein diagnostic accuracy rates of 99.51% and 100% were achieved, respectively, indicating its great potential as a backbone for HD fault diagnosis.