Feature learning for bearing prognostics: A comprehensive review of machine/deep learning methods, challenges, and opportunities

• Proposes a taxonomy for feature learning in bearing prognostics, including temporal, spatial, and spatiotemporal methods. • Details imbalanced data handling techniques, including shallow and adversarial methods. • Reviews feature-invariant learning methods to address inconsistent data distributions across varying operational conditions. • Reviews fusion models with integrated feature learning and prediction stages. • Analyzes bearing benchmark datasets, experimental setups, challenges, and proposes future research directions. Mechanical bearings are common elements in a wide range of applications, such as wind turbines and manufacturing. Therefore, bearing prognostics are crucial to preventing catastrophic failures and machinery breakdowns. In this context, extracting the influential features is often the most challenging task in the prognosis process. This complexity arises because of the non-linear and non-stationary nature of the acquired vibration signals. Therefore, this paper offers an extensive examination of state-of-the-art feature-learning methods. Initially, the paper introduces a taxonomy of feature learning methods, encompassing both shallow and deep learning approaches. The paper also discusses methods of feature-learning under imbalanced data samples and different operational settings. Furthermore, the paper details the experimental setups of commonly used benchmark datasets to assist scholars and practitioners in understanding the subject area. Finally, the study discusses the challenges associated with calculating bearings’ RUL and suggests potential areas for further research.