Fault detection for autonomous underwater vehicles based on interval analysis of feedforward neural networks

This paper proposes an interval analysis method for feedforward neural networks and applies it to the fault detection of autonomous underwater vehicles subject to unknown-but-bounded uncertainties. First, an interval analysis approach is developed to compute the output envelope of a given feedforward neural network trained on fault-free data. Then, by taking measurement noise and approximation errors into account, an adaptive fault detection threshold is derived. Based on this threshold, fault detection is carried out using consistency checking. Compared with conventional semidefinite programming methods, the proposed approach reduces both computational complexity and interval conservatism. Finally, pool experiments conducted on the “Bever-II” AUV validate the effectiveness and applicability of the proposed method.