STCTNet: A Spatio-Temporal Cross-Transformer Network for Autonomous Underwater Vehicle Fault Diagnosis

Autonomous underwater vehicles (AUVs) are critical for marine exploration, making reliable fault diagnosis essential for safe and efficient operation. Modern AUV systems generate heterogeneous multivariate time-series data. However, existing data-driven methods often focus on temporal modeling or singlesignal analysis, failing to fully exploit spatio-temporal dependencies and inter-variable coupling relationships. To address this limitation, a Spatio-Temporal Cross Transformer Network (STCTNet) is proposed for AUV fault diagnosis. The framework adopts a multi-stage architecture to progressively refine feature representations. A Spatio-Temporal Cross Transformer (STCT) module is designed to jointly capture long-range temporal dependencies and dynamic inter-variable interactions, while an Enhanced Attention Fusion (EAF) module adaptively integrates temporal and spatial features. Experimental results demonstrate the superiority of STCTNet, achieving accuracies of 99.18% and 90.02% on the 'Haizhe' and 'Yongyi' AUV datasets, respectively, highlighting the effectiveness of joint spatio-temporal modeling for robust fault diagnosis.