A Lightweight Heterogeneous Neural Network Architecture for Multiscale Feature Extraction in Bearing Fault Diagnosis

Bearings are critical components in rotating machinery, and efficient fault detection is essential for maintaining motor reliability and operational integrity. This study proposes a novel neural network architecture for efficient bearing fault diagnosis that simultaneously captures local and global feature representations from log-Mel spectrograms derived from wideband acoustic emission signals. The framework combines a multiscale convolutional neural network with a Mamba-based selective state space model to improve the discrimination of complex fault patterns while preserving low computational complexity. To support context-aware learning, an adaptive feature fusion strategy is adopted to dynamically integrate features from preceding layers. The proposed framework is further implemented on an embedded platform to evaluate its suitability for edge computing. Experimental results demonstrate that the method achieves superior accuracy and efficiency across multiple fault types and rotational speeds, outperforming existing classification approaches. These results support the development of fully automated bearing condition monitoring systems and underscore the potential of in-sensor computing.