Abstract
Rolling bearing fault diagnosis under complex operating conditions requires discriminative representations of non-stationary vibration signals and maintenance decisions that account for asymmetric economic losses caused by false alarms and missed detections. This paper develops an interpretable and cost-aware diagnostic framework that integrates physical prior knowledge with deep temporal representation learning and maps multiclass diagnostic outputs to maintenance-triggering decisions. The proposed framework adopts a physics-prior-guided dual-channel hybrid feature learning architecture. In the physical feature channel, wavelet packet energy ratios, kurtosis, skewness, and crest factor are extracted as mechanism-related descriptors, while in the deep spatiotemporal feature channel, a one-dimensional convolutional neural network and a bidirectional gated recurrent unit are used to learn high-dimensional temporal representations from normalized vibration segments. The main methodological novelty is a cross-modal multi-head attention module that uses embedded physical features as the query vector to guide the dynamic selection of deep temporal features. A scalar gated fusion mechanism is then introduced to adaptively balance physical robustness and data-driven representation capacity. A posterior-probability-based minimum expected-cost evaluation model is further established to convert diagnostic outputs into binary maintenance decisions under asymmetric false-alarm and missed-detection costs. This economic evaluation component translates diagnostic errors into quantifiable maintenance-cost indicators, providing a unified basis for comparing the economic consequences of downtime losses, redundant inspections, and risks associated with missed faults across different models. The framework is intended for intelligent condition monitoring and risk-sensitive maintenance of rotating machinery, where missed faults and unnecessary maintenance actions may lead to different operational and economic consequences. Experiments on the CWRU, SEU, and in-house rolling bearing test-rig datasets under a unified non-overlapping evaluation protocol show that the proposed framework achieves diagnostic accuracies of 99.88%, 99.75%, and 99.79%, respectively. Ablation studies and feature-distribution analyses show that attention-guided gated fusion improves class separability and reduces maintenance-related errors compared with single-channel models and static fusion strategies. Noise experiments on the CWRU dataset indicate better performance retention than representative data-driven models at 10-20 dB and a relative advantage at 5 dB, although the accuracy decreases markedly under extremely low-SNR conditions such as 0 dB and - 5 dB. Small-sample experiments on the in-house test-rig dataset show higher Accuracy and Macro-F1 than purely data-driven baselines when only 5% and 10% of the training samples are used. Under high missed-detection cost settings, the proposed framework reduces the missed-detection probability of critical faults and lowers the normalized expected maintenance cost. These findings indicate that the proposed method not only improves diagnostic performance but also supports more efficient maintenance resource allocation and reduces expected economic losses in risk-sensitive scenarios, providing an interpretable diagnostic and decision-support tool for economically sensitive industrial maintenance.