A few-shot fault diagnosis method for AUVs based on spatiotemporal representation-change meta-learning

Fault diagnosis methods leveraging multivariate time-series data from sensors can enhance the safety and reliability of AUVs (Autonomous Underwater Vehicles). In practical engineering applications, AUV fault samples are predominantly characterized by few-shot scenarios. However, traditional machine learning techniques require a substantial quantity of labeled fault data samples, rendering them incompatible with the data-scarce nature of AUV fault diagnosis. To the best of our knowledge, this paper is the first to adopt the N-way K-shot task setting for defining the few-shot fault diagnosis problem for AUV multisensor time-series data. We also conduct data preprocessing and partitioning on the publicly available Haizhe AUV dataset to build a benchmark. A few-shot fault diagnosis method is proposed for this problem, designated as SRCML (Spatiotemporal Representation-Change Meta-learning). Further, an in-depth analysis is performed for the spatiotemporal representation-change mechanism. The performance of the method is rigorously validated: the publicly available Haizhe AUV dataset verifies its basic effectiveness, the real-world marine voyage dataset from the Qianlong-2 AUV validates its adaptability to practical underwater scenarios, and the UCI HAR dataset confirms its generalizability.