Unsupervised transfer learning for real-time motor resonance fault diagnosis on programmable logic controllers

Diagnosing motor resonance faults in real-time on programmable logic controllers are challenging due to domain shifts across changing conditions and stringent hardware limits.This study introduces a deep causal adversarial migration transfer learning framework.It synergises multi-physics signal fusion with a physics-guided attention mechanism to disentangle invariant fault features from domain-sensitive variations, followed by adversarial domain alignment.The framework is subsequently lightweighted via structured pruning and quantisation for edge execution.Evaluations on the Case Western Reserve University dataset show the method achieves an average accuracy of 96.2% across varying load tasks, outperforming strong baselines by 3.1%.The final model attains a 31-millisecond inference time, which strictly complies with the sub-100 ms real-time requirement for industrial controllers, proving its effectiveness for dependable edge-based diagnosis.