Physics-Guided Uncertainty Quantification for Trustworthy Bearing Fault Diagnosis

Abstract

Bearing faults in induction motors cause unplanned downtime in continuous manufacturing sectors, where unpre dicted motor failure propagates directly into production delays. Existing deep learning diagnostic models optimize for accuracy but provide no uncertainty estimate, making it impossible to distinguish a confident correct prediction from a confident wrong one. We present PhysicsAE, combining three components: (1) a physics-guided adaptive frequency mask derived from SKF 6205 2RS bearing geometry that reduces autoencoder validation loss by 55.9% over a standard MSE baseline at identical model capac ity; (2) MC Dropout uncertainty quantification producing per prediction confidence scores without additional parameters; and (3) the Physics-Uncertainty Triage (PUT) Framework, a three zone deployment model. Evaluated on the CWRU benchmark, the PUT Framework achieves Zone 3 (confident-wrong)= 0% across all threshold settings in cross-load evaluation and reduces Zone 3 from 41.3% to 39.6% in cross-speed evaluation. Five-seed vali dation confirms the fault-severity uncertainty ordering holds in 5/5 seeds with standard deviation <0.0001. An ablation confirms that physics-guided training and encoder-embedded Dropout are jointly necessary: removing either collapses uncertainty to zero.