Research Review and Future Prospects of Intelligent Fault Diagnosis Technology for Rolling Bearings

Abstract

Intelligent fault diagnosis for rolling bearings is central to achieving predictive maintenance in industrial equipment. Traditional methods rely on manual feature engineering and exhibit weak generalization under complex operating conditions. Deep learning, through end-to-end learning, can automatically extract fault features from raw data, driving a paradigm shift in diagnostic approaches. Current mainstream methods include deep temporal models that directly process one-dimensional vibration signals (e.g., 1D-CNN) and visual models based on time-frequency images (e.g., Vision Transformer). To address challenges such as data scarcity and varying working conditions, techniques like transfer learning and generative adversarial networks have been widely applied. However, key bottlenecks for practical engineering implementation remain, including poor model interpretability, insufficient robustness under strong noise, weak generalization with small samples, and difficulties in lifetime prediction. Future research should focus on explainable artificial intelligence, digital twin-based data augmentation, physics-informed data fusion, and lightweight edge deployment to build more trustworthy, adaptive, and forward-looking intelligent diagnostic systems.