Ensemble-Based Uncertainty Estimation for Fault Diagnosis in Motor Drive Systems

This study proposes a supervised-learning-based fault diagnosis framework for motor drive systems, enhanced by a ensemble-based uncertainty estimation technique. Although supervised diagnostic models typically achieve strong performance for known fault patterns, their prediction reliability degrades when facing operating conditions or fault types not included in the training data. In addition, certain fault classes inherently exhibit low discriminability, leading to unstable predictions and reduced diagnostic accuracy. To address these limitations, the proposed framework quantifies epistemic uncertainty to directly assess the confidence of each diagnostic output. By analyzing how uncertainty behaves for poorly learned classes, misclassified samples, and previously unseen fault scenarios, the results demonstrate that uncertainty measures provide strong discriminative power between in-distribution and out-of-distribution inputs. This capability enables effective identification of unreliable predictions and contributes to improving the practical robustness and trustworthiness of data-driven motor-drive fault diagnosis systems.