Few-Shot Fault Diagnosis of Harmonic Reducer of Industrial Robot Based on TCIFMN

摘要

Harmonic reducers (HRs), as precision components in industrial robots, are susceptible to failure in complex working conditions. The current intelligent diagnosis methods for HR are limited and lack sufficient fault data to train deep learning models effectively. This article presents a novel few-shot learning (FSL) fault diagnosis model named transformer convolution integration feature metric network (TCIFMN) which is based on metric learning (ML) and combines the advantages of convolutional neural networks (CNNs) module and transformer architecture. The transformer convolution integration (TCI) module enhances the capability of extracting fault information and provides an input pair for feature metric network (FMN) which calculates similarity scores (SMLS) based on the idea of the relation network (RN). In addition, to enhance the fault features of HR, the continuous wavelet transform (CWT) is used to turn the fault signal into images in this article. CWT can represent the fault information in the time and frequency domains. The experimental results demonstrate that even with extremely limited training data, TCIFMN achieves prominent performance with 93% accuracy under complex working conditions, surpassing other comparative models.