Abstract
Bearings are among the most failure-prone components in rotating systems, making early fault detection crucial in industrial applications. While recent publications have focused on this issue, challenges remain, particularly in dealing with heavy-tailed or non-cyclic impulsive noise in recorded signals. Such noise poses significant challenges for classical fault selectors like kurtosis-based methods. Moreover, many deep-learning approaches struggle in these environments, as they often assume Gaussian or stationary noise and rely on large labeled datasets that are rarely available in practice. To address this, we propose a robust sparse non-negative matrix factorization (NMF) method based on the maximum-correntropy criterion, which is known for its robustness in the presence of heavy-tailed noise. This methodology is applied to identify fault frequency bands in the spectrogram of the signal. The effectiveness of the approach is validated using simulated fault signals under both Gaussian and heavy-tailed noise conditions through Monte Carlo simulations. A statistical efficiency analysis confirms robustness to random perturbations. Additionally, three real datasets are used to evaluate the performance of the proposed method. Results from both simulations and real-world data demonstrate the effectiveness of the proposed approach.