A Fault Diagnosis Method for AUVs Based on Meta-Self-Attentive Variable-Scale CNN

Ocean investigation is made possible by autonomous underwater vehicles (AUVs), and the safe navigation of these vehicles depends on the ability to diagnose actuator faults quickly. However, the complicated marine environment and the restricted availability of failure data due to unexpected failures present a significant difficulty for troubleshooting. A meta-self-attention variable-scale convolutional neural network (MSAVS-CNN) model for fault diagnosis of AUVs is proposed in this paper. The model directly utilizes raw vibration data collected from sensors as input. To enhance the convergence speed, a subtask-based gradient optimization method is employed during model fitting. In the feature extraction process, a self-attentive variable-scale approach is employed, enabling the acquisition of information at different scales and facilitating the autonomous learning of crucial features through convolutional kernel size adjustments. In situations where sample availability is limited, the suggested method leverages a meta-learning approach to train the diagnostic model, thereby improving its ability to generalize and enabling cross-domain diagnosis. Experiments confirm the validity of the method and show that the suggested method can diagnosis the actuator failure of AUVs with few samples.