Cost-sensitive active learning for rotating machinery fault diagnosis under data imbalance

Data-driven fault diagnosis has become an important approach for ensuring the reliability and safety of rotating machinery. However, practical deployment is often hindered by severe data imbalance and label scarcity. Active learning (AL) mitigates label scarcity by selecting the most informative samples for annotation to improve model performance. Nevertheless, conventional AL strategies typically optimize for overall accuracy and treat all misclassifications equally. Consequently, they fail to account for the asymmetric operational costs of diagnostic errors, particularly the greater losses caused by missed detections and severity underestimation. To bridge this gap, this paper proposes a cost-sensitive active learning (CS-AL) framework tailored for highly imbalanced fault distributions. In this framework, raw vibration signals are first transformed into two-dimensional time–frequency representations via continuous wavelet transform (CWT). A Bayesian convolutional neural network, integrated with a convolutional block attention module, is developed to extract fault-sensitive features, while epistemic uncertainty is quantified through Monte Carlo (MC) dropout. To explicitly penalize high-risk errors during sample selection, an asymmetric diagnostic cost matrix is constructed to formulate a cost-sensitive acquisition score based on the expected pairwise misclassification cost. To optimize annotation efficiency, a diversity-aware clustering mechanism selects representative prototypes from this high-risk candidate pool. Experimental validations on the CWRU bearing and MCC5-THU gearbox datasets demonstrate that the CS-AL framework achieves enhanced diagnostic performance and significantly reduces high-cost misclassification risks compared to representative AL baselines. The results confirm the effectiveness of CS-AL in accurately allocating limited labeling resources to rare, safety-critical faults while minimizing total operational risk under imbalanced conditions.