Deep Reinforcement Learning Approach for X-rudder AUVs Fault Diagnosis Based on Deep Q-network

Abstract

The rudder mechanism of the X-rudder autonomous underwater cehicle (AUV) is relatively complex, and fault diagnosis capability is an important guarantee for its task execution in complex underwater environments. However, traditional fault diagnosis methods currently rely on prior knowledge and expert experience, and lack accuracy. In order to improve the autonomy and accuracy of fault diagnosis methods, and overcome the shortcomings of traditional algorithms, this paper proposes an X-steering AUV fault diagnosis model based on the deep reinforcement learning deep Q network (DQN) algorithm, which can learn the relationship between state data and fault types, map raw residual data to corresponding fault patterns, and achieve end-to-end mapping. In addition, to solve the problem of few X-steering fault sample data, Dropout technology is introduced during the model training phase to improve the performance of the DQN algorithm. Experimental results show that the proposed model has improved the convergence speed and comprehensive performance indicators compared to the unimproved DQN algorithm, with precision, recall, F _1−score, and accuracy reaching up to 100%, 98.07%, 99.02%, and 98.50% respectively, and the model’s accuracy is higher than other machine learning algorithms like back propagation, support vector machine.

RIGHTS & PERMISSIONS

Harbin Engineering University and Springer-Verlag GmbH Germany, part of Springer Nature