A fault diagnosis method with multi-source data fusion based on hierarchical attention for AUV. (15th December 2022)

• Record Type: Journal Article
• Authors: Xia, Shaoxuan; Zhou, Xiaofeng; Shi, Haibo; Li, Shuai; Xu, Chunhui
• Is Part Of: Ocean engineering. Volume 266(2022) Part 1
• Publication Date: 2022-12-15
• Subjects: AUV fault diagnosis -- Multi-source data fusion -- Hierarchical attention -- Random mask -- Feature orthogonalization; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2022.112595
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: Multi-source data fusion is an important method to improve the performance of Autonomous Underwater Vehicle (AUV) fault diagnosis. However, most of the current fault diagnosis methods are based on a single data source or treat multi-source data as single. Firstly, we demonstrate the necessity of multi-source data fusion and propose a universal data hierarchy. Then, a hierarchical attention based multi-source data fusion method is proposed for fault diagnosis (HAMFD). The method consists of an encoder–decoder network, a fusion network stacked with encoders and attention mechanisms, and a fault recognition method based on attention distribution. The fusion network uses the encoder and hierarchical attention to extract the deep features, and fuse the features hierarchically. We use the multi-layer attention distribution to explain the fault evaluation and realize fault recognition. A random mask fusion strategy is designed for redundancy and a feature orthogonalization method is proposed for the strong coupling among multiple data sources. The proposed method is validated on the monitoring data of Qianlong-2 AUV obtained during the sea trial in the South China Sea. The fault detection rate is more than 98%, the recognition rate is about 100% for strong faults, and more than 90% for other faults. Highlights: For multi-source of AUV data, hierarchical attention is applied for data fusion. A universal four-layer hierarchy of AUV multi-source data is proposed. Fault recognition through the interpretability of attention mechanism. We proposed feature orthogonalization and random mask for the redundancy. The experiments on Qianlong-2 AUV show effectiveness and superiority.