A modified transformer-based fault diagnosis method for robotic motor systems

Robotic motors, particularly permanent magnet synchronous motors (PMSMs), often experience multiple simultaneous faults, which limits the effectiveness of traditional single-fault diagnostic methods. This paper proposes a hybrid fault diagnosis approach that combines simulated and historical data to improve classification accuracy. The pre-processed current signals are deeply feature-mined by a convolutional neural network (CNN), and the timing correlation properties are subsequently resolved by a feature enhancement module integrating an improved Transformer structure. Comparative experimental analyses show that the fusion architecture significantly improves the recognition accuracy compared with traditional methods in the multi-fault classification task, which validates the engineering applicability of the proposed framework under complex working conditions.