A Risk-Sensitive Fault Diagnosis Framework for AUVs Using Learnable Dual-Branch Bandpass Filtering and Bayesian Minimum-Risk Decision

An autonomous underwater vehicle (AUV) is an important operation platform in ocean exploration engineering, and the operation state of its actuator directly determines the navigation safety. However, the marine environment is complex and changeable, and the state signal characteristics of autonomous underwater vehicles are complex and vulnerable to noise interference. The performance of the traditional fault identification model degrades significantly in the noise environment, and the model ignores the risk difference of misjudgment, which may lead to serious misjudgment consequences. In order to solve the above problems, we propose a Risk-Sensitive Dual-Branch Learnable Filtering (RS-DBLF) framework for AUV fault diagnosis. The feature extraction backbone network adopts a dual-branch structure, which can adaptively capture the low-frequency slow degradation fault features and high-frequency impact fault features, and simultaneously complete the noise suppression and multi-scale fault frequency domain feature separation. Secondly, a fault asymmetric cost matrix is established for the cost-sensitive diagnosis scenario, and the expected misjudgment loss corresponding to various faults is calculated by using the posterior probability of the network output. The global minimum-risk criterion is used to replace the traditional cross entropy hard classification reasoning, so as to reduce the high-cost fault misjudgment probability from the decision level. The “Haizhe” open dataset is used to verify the effectiveness of the algorithm. The results show that the cost-sensitive dual-branch learning bandpass network has a stronger robustness in noise environments, and the minimum-risk decision based on Bayesian can effectively reduce the loss of misjudgment.