Autonomous Underwater Vehicles Key Components Fault Diagnosis Method Based on Multisource Information Multilevel Fusion With Fault Feature Exchange

In order to enhance the accuracy and stability of fault diagnosis in autonomous underwater vehicles (AUV), this paper proposes an AUV fault diagnosis method based on multisource information multilevel fusion with fault feature exchange. This method consists of three main components. First, multiple convolutional neural networks (CNN) are used to extract and analyze the fault features contained in the multisource signals to obtain the judgment of each signal source for various types of faults, namely the credibility value of the evidence body. Second, an fusion module is designed innovatively to share and combine generic fault features from multiple signal sources, facilitating mutual correction and supplementation of fault features from different branches, thus improving comprehensive fault recognition, and the fusion module is trained using a meta‐learning approach, where reinforcement learning (policy gradient) is incorporated to optimize the controller and adaptively obtain the optimal fusion coefficients, thereby enhancing the adaptability and performance of the model. Finally, the weighted Dempster‐Shafer evidence theory is proposed to fuse the belief values of various pieces of evidence to obtain the final decision result. Experimental results demonstrate that the proposed method achieves high precision and stability in AUV fault detection and exhibits strong fault tolerance in the presence of data missing.