Out-of-distribution detection for vibration-based fault diagnosis using generalized extreme value theory

Abstract

Current deep learning fault diagnosis networks for mechanical equipment may fail to reliably identify out-of-distribution (OOD) faults when samples from previously unseen fault categories are unavailable during model development, leading to the misclassification of such OOD samples as known categories. To address this issue, a deep learning fault diagnosis framework based on the generalized extreme value theory (GEVT) is proposed. First, this framework replaces the traditional label with one sample, providing waveform-level evidence for the resulting decisions, enabling users to intuitively confirm results during both training and testing. Then, classification is performed based on the mean absolute distance (MAD) between network predictions and sample labels to enhance interpretability. Finally, the GEVT is employed to fit the extreme values of the MAD, and the cumulative distribution function probability is calculated. Based on this probability, a threshold is determined to detect OOD faults. Validation on the PHM2009 gearbox dataset and BJTU-RAO bogie dataset demonstrates that this framework not only effectively identifies previously unseen OOD samples but also maintains high detection accuracy for in-distribution (InD) data.