Improved state-χ2 fault detection of Navigation Systems based on neural network

In INS/GPS Integrated Navigation Systems, the classic state-χ² testing method is used to ascertain if any fault exists by comparing a priori information with measurement results and examining whether the structure of the mean and covariance matrix of the n-DOF of Gaussian distributed random vector is consistent with the hypothetic values. A fault can be found with this method; however, it fails to tell the fault exists whether in the INS system or in the GPS part. This paper presents an improved neural network-based residual χ² testing technique to solve this problem; i.e., the output of the trained neural network is substituted for the INS system output when a fault is detected at first time, and the state-χ² testing algorithm is resumed. The simulation results show whether the fault comes from the INS system or the GPS system. Simulation experiments demonstrate its feasibility.