LSBT-Net: A lightweight framework for fault diagnosis of bearings based on an interpretable spatial-temporal model
Yicheng Duan, Tongguang Yang, Chenlin Wang, Yongjian Zhang, Qingkai Han, Shuangping Guo
Abstract
Intelligent fault diagnosis based on deep learning has emerged as a research focus in mechanical equipment due to its adaptive feature extraction capability. However, current models struggle with low accuracy, high computational costs, and poor interpretability when detecting faults in insulated bearings. To address these challenges, this paper proposes a novel lightweight spatiotemporal model-based intelligent diagnostic framework, named LSBT-Net, which aims to identify motor insulating bearing faults in practical engineering applications more accurately. Specifically, this research breaks the conventional thinking of “learning fault data feature information” by innovatively developing a spatiotemporal information fusion module. This module is cleverly integrated into the LSBT-Net framework, enabling the extraction of both local and global high-dimensional fault feature information from insulating bearings. At the same time, based on a lightweight design, it significantly reduces the total number of parameters and computational resources required by the framework, thus lowering its computational complexity. The t-SNE algorithm is introduced into the LSBT-Net framework to achieve local or global interpretability. Furthermore, by calculating the gradient information of the LSBT-Net framework on the fault types of insulating bearings through backpropagation , the interpretability of the framework with respect to the physical information is enhanced. Using insulating bearings and typical fault experiments as examples, the LSBT-Net framework demonstrates excellent diagnostic capability and generalization performance compared to other advanced methods.