A multi-scale spatial–temporal capsule network based on sequence encoding for bearing fault diagnosis

• Published in: Complex & intelligent systems Vol. 10; no. 5; pp. 6189 - 6212
• Main Authors: Wang, Youming, Chen, Lisha
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.10.2024; Springer Nature B.V; Springer
• Subjects: Algorithms; Attention mechanism; Automation; Bearing fault; Capsule network; Coding; Complexity; Computational Intelligence; Data Structures and Information Theory; Dilated convolution; Engineering; Fault diagnosis; Fault tolerance; Intelligent systems; Machine learning; Machinery; Multi-scale; Neural networks; Original Article; Time series; Attention mechanism; Dilated convolution; Multi-scale; Bearing fault; Capsule network
• ISSN: 2199-4536; 2198-6053; 2198-6053
• DOI: 10.1007/s40747-024-01462-8

The Capsule Network (CapsNet) has been shown to have significant advantages in improving the accuracy of bearing fault identification. Nevertheless, the CapsNet faces challenges in identifying the type of bearing fault under nonstationary and noisy conditions. These challenges arise from the distinctive nature of its dynamic routing algorithm and the use of fixed single-scale kernels. To address these challenges, a multi-scale spatial–temporal capsule network (MSCN) based on sequence encoding is proposed for bearing fault identification under nonstationary and noisy environments. A spatial–temporal sequence encoding module focuses on feature correlations at various times and positions. Dilated convolution-based multiscale capsule layer (MCaps) is designed to capture spatial–temporal features at different scales. MCaps establishes connections between various layers, enhancing the comprehension and interpretation of spatial–temporal features. Furthermore, the Bhattacharyya coefficient is introduced into the dynamic routing to compare the similarity between capsules. The validity of the model is verified through comparative experiments, and the results show that MSCN has significant advantages over traditional methods.