DAE–DBN model for continuous thruster fault identification in underwater vehicles

Abstract Thrusters are critical components affecting the maneuvering accuracy and operational reliability of underwater vehicles; therefore, precise fault diagnosis is essential to ensure safe operation and effective fault-tolerant control. However, existing methods are limited in continuous fault severity identification, generalization to unseen fault modes, and disturbance rejection under measurement noise. In this paper, a denoising autoencoder–deep belief network (DAE–DBN) framework is proposed for thruster fault identification. First, denoising and feature extraction are performed on raw multi-channel signals using a DAE. The extracted features are subsequently fed into a DBN to achieve continuous fault severity identification via regression. Experiments are conducted in a controlled pool environment considering six fault scenarios and one normal-operation scenario. Comparative studies with classical regression approaches, self-organizing map, backpropagation neural network, and a standalone DBN are performed. The results demonstrate high identification accuracy on training samples, reliable generalization to unseen fault conditions, and superior disturbance rejection capability of the proposed DAE–DBN model.