An improved hypergraph convolutional network based on multi-channel fusion signals for semi-supervised fault diagnosis of autonomous underwater vehicle thrusters

Abstract Autonomous underwater vehicle (AUV), as a highly efficient tool for ocean exploration, relies on thrusters whose fault diagnosis is a key aspect to ensure safe navigation. However, single-channel signals cannot fully capture the complex fault information of AUV thrusters, and the process of obtaining labeled fault samples also faces challenges of high difficulty and high cost. To address them, this paper proposes a semi-supervised approach for diagnosing faults in AUV thrusters fusion signals based on the improved hypergraph convolutional network (HGCN). By adding the auxiliary of the noise-channel signal, the noise-assisted multivariate empirical modal decomposition algorithm is first used to process the original dual-channel vibration signals to generate a new fusion signal. Subsequently, the fusion signals are constructed into a hypergraph structure based on the k -nearest neighbor algorithm, and the HGCN model is improved by embedding the graph attention and gate recurrent unit modules, thereby yielding more robust and expressive node representations. Finally, the proposed method is comprehensively validated using the measured dataset from the Haiwei No.1 AUV of the 713 Research Institute in China. The experimental results indicate that the multi-channel fusion signals greatly enhance the fault diagnosis performance of AUV thrusters, the improved modules of HGCN also produce significant effects, and the improved HGCN model outperforms other state-of-the-art graph deep learning models and classical non-graph deep learning models under various low-labeling-rate scenarios.