Hybrid feature adaptive fusion network for multivariate time series classification with application in AUV fault detection

Abstract

ABSTRACT Autonomous underwater vehicles (AUVs) acquire large-scale multivariate time series (MTS) data during navigation, which can be utilised to realise fault diagnosis, condition monitoring, and other functions by means of classifying the monitoring data. However, due to the complexity and time-variation of relationships between many variables of the MTS, we propose a MTS classification method, namely hybrid feature adaptive fusion network (HFAF). Specifically, a multi-scale method is first proposed to generate monitoring windows with different scales, and the spatiotemporal information is then fully obtained by dilated convolutional neural network (D-CNN) and dilated recurrent neural network (D-RNN). Subsequently, an adaptive feature fusion network based on an attention mechanism is introduced to address the redundancy and conflict between different scales. Finally, the hybrid feature network and adaptive fusion network are stacked up to form HFAF. The effectiveness and superiority of HFAF in AUV fault detection are demonstrated by the experiments conducted on Haizhe AUV, which yields more than 96% precision and more than 95% recall for various faults, outperforming other fault detection methods.