A wavelet-enhanced multiscale dense capsule network for rotating machinery fault diagnosis

• Published in: Journal of mechanical science and technology Vol. 39; no. 10; pp. 5763 - 5776
• Main Authors: Li, Yongjian, Gan, Yi, Wu, Yuyuan, Chen, Zhaoyang, Lv, Qiuxia, Xiong, Qing
• Format: Journal Article
• Language: English
• Published: Seoul Korean Society of Mechanical Engineers 01.10.2025; Springer Nature B.V; 대한기계학회
• Subjects: Control; Dynamical Systems; Engineering; Fault diagnosis; Feature extraction; Industrial and Production Engineering; Mechanical Engineering; Original Article; Robustness; Roller bearings; Rotating machinery; Vibration; Wavelet analysis; 기계공학; Fault diagnosis; Dilation convolution; Capsule network; Wavelet convolutional; Dense connectivity
• ISSN: 1738-494X; 1976-3824
• DOI: 10.1007/s12206-025-0914-x

Owing to its sensitivity to spatial details, the capsule network (CapsNet) has demonstrated superior performance in fault diagnosis. However, CapsNet is typically restricted to a single scale and shallow architecture in the time domain, limiting its ability to represent features and reducing its robustness under complex vibration data and noisy conditions. Therefore, a wavelet-enhanced multiscale dense capsule network (WMDCN) is proposed, offering improved multi-scaled feature representation and a robust, dense structure. The SE-wavelet (SEW) feature extractor integrates wavelet domain analysis with dynamic adjustments, adaptively extracting critical time-frequency characteristics. The multiscale dense connection block (MSDCB) employs dilated convolutions and dense connections to integrate multiscale features and capture deeper information, enhancing feature propagation. Hellinger distance (HD) is incorporated into the dynamic routing algorithm to provide a comprehensive nonlinear similarity between capsule distributions, refining feature aggregation. Finally, the proposed WMDCN is effectively applied to fault identification, utilizing the learned features while preserving spatial correlations. The diagnostic performance and noise robustness of the model are validated on two rolling bearing datasets, outperforming traditional approaches.