Aiming at the problem that the low efficiency of detecting the soft faults by the traditional chi-square test method in the fault diagnosis of loosely-coupled inertial/satellite integrated navigation systems, a double-threshold test method to detect soft faults assisted by BP neural network is proposed. The location and velocity sub-predator is established by BP neural network, which can realize the prediction of satellite measurement data. On this basis, a double-threshold test is proposed according to the prediction accuracy, which can assist the residual chi-square component detection method to detect faults and reconstruct the system. The simulation results show that the proposed method can effectively improve the filtering accuracy when fault occurs, and reduce the missed alarm rate and improve the reliability of integrated navigation.