Intelligent fault diagnosis of rotating machinery under complex working conditions: a condensed survey

Abstract

Rotating machinery fault diagnosis under complex working conditions is central to predictive maintenance, yet variable speed, load fluctuation, environmental noise and compound faults often break the distributional assumptions of conventional supervised models. This revised condensed survey strengthens the comparison among deep learning, transfer learning and physics-informed fusion by explicitly contrasting their advantages, limitations, applicable scenarios and deployment risks. Deep learning offers strong feature representation and can reach about 97% accuracy under controlled conditions, but it commonly loses 8-14 percentage points when operating conditions become complex. Transfer learning reduces cross-domain gaps through statistical alignment, adversarial adaptation, few-shot learning and multi-source domain adaptation, while physics-informed fusion embeds mechanical constraints and signal priors to improve out-ofdistribution robustness. A new loss-function explanation is added to clarify how fault-frequency relationships, dynamic residuals and energy-consistency terms can be incorporated into the training objective. The synthesis indicates that future fault diagnosis systems should integrate data-driven representation learning, cross-domain adaptation and physically meaningful inductive biases while reporting uncertainty and deployment limitations.