A New Multisensor Feature Fusion KAN Network for Autonomous Underwater Vehicle Fault Diagnosis

Autonomous underwater vehicle (AUV) plays a pivotal role in ocean exploration. The failure of AUV can directly affect its efficiency and safety, leading to extensive research on AUV fault diagnosis. However, the utilization of a single data-source-driven approach for AUV fault diagnosis fails to comprehensively depict the comprehensive fault states of AUV as intelligent electromechanical devices, which poses significant challenges for AUV fault diagnosis. To address this challenge, a new multisensor feature fusion KAN network (MFKAN) is proposed. First, a multisensor faults enhancement module is proposed, aimed at exploiting the fault expression potential of multisensor data. Second, a multisensor feature fusion network based on a convolutional neural network is designed to extract global fault features and local correlation features of AUV, achieving feature-level integration of fault features. Third, by combining the effective-Kolmogorov-Arnold Networks (KANs) with the multisensor feature fusion network, an MFKAN is constructed to identify diverse fault states of AUV. MFKAN demonstrated superior performance on the "Haizhe" AUV dataset. The results show that MFKAN has achieved an accuracy rate of 99.18% in AUV fault diagnosis, highlighting its exceptional fault diagnosis performance.