Evaluating deep learning models for fault diagnosis of a rotating machinery with epistemic and aleatoric uncertainty
A new research paper published on arXiv explores the effectiveness of various deep learning models in diagnosing faults in rotating machinery, specifically focusing on their ability to handle uncertainty. The study compares models like dropout-based methods, Bayesian neural networks, and deep ensembles, evaluating their performance in detecting out-of-distribution data caused by unseen faults (epistemic uncertainty) and noise (aleatoric uncertainty). Findings indicate that deep ensemble models generally outperform others in detecting epistemic uncertainty, while also showing more robust performance against aleatoric uncertainty, making them a preferred choice due to their accuracy and efficiency. AI
IMPACT Provides guidance on selecting uncertainty-aware deep learning models for reliable fault diagnosis in industrial machinery.