Actuator Fault Reconstruction via Dynamic Neural Networks for an Autonomous Underwater Vehicle Model

This paper proposes the development of a scheme for the fault diagnosis of the actuators of a simulated model accurately representing the behaviour of an autonomous underwater vehicle. The Fossen model usually adopted to describe the dynamics of the underwater vehicle has been generalised in this paper to take into account time-varying sea currents. The proposed fault detection and isolation strategy uses a data-driven approach relying on multi-layer perceptron neural networks that include auto-regressive exogenous prototypes that provide the fault reconstruction. These tools are thus exploited to design a bank of dynamic neural networks for residual generation that are trained on the basis of the input and outputmeasurements acquired from the simulator. In this work, the residuals are designed to represent the reconstruction of the fault signals themselves. Moreover, the neural network bank is also able to perform the isolation task, in case of simultaneous and concurrent faults affecting the actuators. The paper firstly describes the steps performed for deriving the proposed fault diagnosis solution. Secondly, the effectiveness of the scheme is demonstrated by means of high-fidelity simulations of a realistic autonomous underwater vehicle, in the presence of faults and marine current.