Noise in Rolling Bearing Fault Diagnosis: From Characterization to Robust Solutions

Abstract

Rolling bearing fault diagnosis is essential for the reliability and safety of rotating machinery, yet its effectiveness in industrial practice is fundamentally challenged by noise. This narrative review argues that noise in bearing fault diagnosis should not be treated as a single additive disturbance, but rather as a collection of distinct phenomena — environmental and background signal noise, sensor and acquisition noise, operating-condition-induced domain shift, label noise, and compound noise scenarios — each affecting different stages of the diagnostic pipeline and demanding differentiated strategies. This review develops a noise taxonomy for rolling bearing fault diagnosis and systematically examines how each noise type degrades signal processing, feature extraction, and diagnostic decision-making. It surveys denoising strategies, noise-robust feature engineering approaches, and noise-aware diagnostic models through this noise-type lens, identifying which strategies are appropriate for which noise conditions. A standardized evaluation and benchmarking protocol is proposed to address the inconsistency that currently prevents meaningful comparison of noise-robustness claims across studies. The central synthetic contribution is a noise–method matching matrix that maps each noise category to appropriate combinations of denoising, feature, model, and evaluation strategies, providing a structured framework for method selection and future research design. The review concludes that robust bearing fault diagnosis cannot be achieved through a single universal model. The more realistic path forward requires noise-type-aware strategies, standardized multi-noise evaluation, and methods co-designed for industrial deployment constraints.