Few-Shot Fault Diagnosis of Harmonic Reducers Used in Industrial Collaborative Robots Using Prototypical and Siamese Networks

Abstract

Harmonic reducers are critical components in industrial collaborative robots and are highly susceptible to fault development due to continuous use in high load and torque conditions. The fault diagnosis of harmonic reducers to ensure their reliability is therefore essential, particularly under the practical constraint of limited fault data. This study proposes a few-shot learning framework for harmonic reducer fault diagnosis using vibration and current signals collected under three fault conditions. Prototypical and Siamese network architectures with feature extraction modules including time-frequency mixing (TFM), convolutional neural networks (CNN), and denoising spatial attention convolutional neural networks (DSACNN), are systematically evaluated for robustness to unseen data and complex operating conditions. Experimental results show that vibration signals provide stronger fault-discriminative information, with TFM-based models achieving the best overall performance, while current-signal-based diagnosis is more sensitive to the choice of feature extractor due to low signal-to-noise ratio, and the CNN-based models give superior classification accuracy in this case. Overall, both the prototypical and Siamese networks perform competitively under unseen operating conditions, confirming their effectiveness for low-data fault diagnosis in harmonic reducers and their strong potential as source models for transfer to a fully operational industrial robot.