A novel hierarchical structural pruning-multiscale feature fusion residual network for intelligent fault diagnosis. (June 2023)

• Record Type: Journal Article
• Authors: Cheng, Yiwei; Lin, Xinnuo; Zhu, Haiping; Wu, Jun; Shi, Haibin; Ding, Huafeng
• Is Part Of: Mechanism and machine theory. Volume 184(2023)
• Publication Date: 2023-06
• Subjects: Intelligent fault diagnosis -- Deep learning -- Multiscale feature fusion residual network -- Hierarchical structural pruning -- Vibration signals; Machine theory -- Periodicals; Machinery -- Periodicals; Machines -- Périodiques; Génie mécanique -- Périodiques; Machine theory; Machinery; Periodicals; 621.81
• DOI: 10.1016/j.mechmachtheory.2023.105292
• Languages: English
• ISSNs: 0094-114X

Abstract

Highlights: HSP-MFFRN is proposed for multi-scale feature extraction, fusion and compression. Bilinear interpolation is utilized to broaden the receptive field of convolution filters. Hierarchical structural pruning is designed for model compression and lightweight. Comprehensive experiments and comparison with existing methods are conducted. Abstract: Residual learning is a commonly used method in the intelligent fault diagnosis (IFD) field. The existing residual networks usually have a large model volume, whose training process puts forward high requirements for both time consumption and the computing resource configuration. This paper proposes a novel hierarchical structural pruning-multiscale feature fusion residual network (HSP-MFFRN) for IFD. The multiple multi-scale feature extraction modules and feature fusion modules are designed in the proposed HSP-MFFRN to extract, fuse and compress the multi-scale features without changing the size of the convolutional filter. In addition, hierarchical structural pruning is implemented for HSP-MFFRN to delete redundant channels. Two experimental study cases are conducted to verify model diagnosis performance, including a rolling bearing IFD case and an aerostat capsule IFD case. The experimental results show that the proposed model can effectively compress the learnable parameters and model volume after hierarchical structural pruning, and can still achieve superior diagnostic performance compared with other deep learning methods, residual learning methods, and other advanced methods reported in the literature.