Angular resampling-sparse representation classification (AR-SRC): A new method for bearing fault diagnosis in non-stationary conditions

Abstract

This paper presents a novel methodology for bearing fault diagnosis under non-stationary operating conditions using angular resampling combined with sparse representation classification. The proposed approach addresses variable speed challenges by transforming time-domain vibration signals into the angular domain through encoder-based resampling, enabling extraction of speed-invariant envelope order spectrum features. For classification, a structured dictionary is constructed from training samples for each bearing condition (healthy, inner race, outer race, ball, and combined faults). Sparse coding is performed using the fast iterative shrinkage-thresholding algorithm (FISTA) within an ℓ₁-norm regularized framework, with final class assignment determined by minimizing reconstruction error across class-specific dictionaries. The proposed framework achieved 99.86% cross-validated accuracy on the Ottawa dataset and demonstrated superiority over state-of-the-art methods (classical machine learning classifiers and some deep learning architecture), all evaluated on identical speed-invariant features. Comprehensive cross-domain validation confirmed robust generalization: cross-speed-profile experiments achieved 99.6% mean accuracy when testing on unseen speed dynamics, while cross-load validation achieved 98.8% mean accuracy across different bearing types and loading conditions. All cross-domain scenarios exceeded 97.8% accuracy. This integrated framework provides a transparent, physically interpretable solution for bearing fault diagnosis, offering enhanced robustness to speed variability and load variations, with practical applicability for industrial condition monitoring under realistic operating conditions.