Energy-Guided Wavelet Activation-Pooling for Intelligent Bearing Fault Diagnosis

Rolling bearings serve as critical components within rotating machinery, and their operational reliability is fundamental to industrial productivity and system safety. Conventional fault diagnosis methods, though mature, often rely on expert knowledge and manual feature engineering, limiting their scalability and adaptability. Deep learning approaches offer end-to-end feature learning capabilities but are susceptible to noise and redundant information, as standard activation and pooling operations are not optimized for non-stationary signals. To mitigate these challenges, this paper introduces a wavelet-driven activation-pooling module tailored for intelligent bearing fault diagnosis. The module first applies discrete wavelet transform to decompose intermediate convolutional features into multiple frequency sub-bands. Subsequently, an energy-aware adaptive activation mechanism is employed to emphasize fault-related frequency components while attenuating bands dominated by noise. Furthermore, multi-scale feature alignment and frequency-band compression operations are incorporated to maintain compatibility with common backbone network architectures, thereby supporting plug-and-play integration without structural redesign. Comprehensive evaluations are conducted on the publicly available XJTU-SY bearing dataset. The experimental results confirm that the proposed module consistently achieves higher diagnostic accuracy and demonstrates stronger robustness across a range of signal-to-noise ratios when compared to conventional activation-pooling schemes. These findings highlight the module's practical potential for noise-resilient fault diagnosis in real industrial environments.