CARENet: a novel interpretable channel attention residual enhanced network for gearbox fault diagnosis under strong noise conditions

Abstract

Gearbox fault diagnosis is crucial for ensuring production safety and improving operational efficiency. However, gearbox fault diagnosis still faces challenges due to strong noise environments. To address these challenges, we propose a novel interpretable channel attention residual enhancement network (CARENet) for robust fault diagnosis under noisy conditions. Vibration signals are first transformed into time-frequency representations using continuous wavelet transform, enhancing the characterization of transient fault features. A feature learning residual enhancement module composed of stacked multi-level channel enhanced attention blocks (CEABlocks) is designed to highlight discriminative feature channels despite noise interference. The CEABlocks capture multi-scale contextual information via dilated convolutions and enhance fault-relevant features through channel attention, while the residual architecture with short-range skip connections facilitates deep feature propagation and improves training stability. Experiments on two gearbox datasets demonstrate that CARENet consistently outperforms existing methods under varying signal-to-noise ratios, achieving highly reliable classification. Gradient-weighted Class Activation Mapping visualizations reveal that the network focuses on physically meaningful fault regions, illustrating its interpretability. In summary, this paper emphasizes its broader relevance to industrial fault monitoring, uncertainty-aware measurement, and multidisciplinary applications in data-driven diagnostics.