Recurrent neural network applied to fault diagnosis of Underwater Robots

Study of thruster fault diagnosis of Underwater Robots (URs) is undertaken to improve its whole system reliability. Based on the BP neural network, an improved recurrent neural network (RNN) is proposed and the network training algorithm is deduced. The RNN is trained by voyage head and yaw turning experiments, and the well trained network is applied to model for the URs. Compared the model's outputs with the sensors' outputs, the residuals can be obtained; Fault detection rules can be distilled from the residuals to execute thruster fault diagnosis. The methods presented here are applied to the simulation and sea trial experiments, and plenty of results are got. Based on the analysis of the experiments results, the validity and feasibility of the methods can be verified, and some reference values in engineering application can be demonstrated by the results.