Aimed at the problems that integrated navigation system is rapid divergent in error when global navigation satellite system (GNSS) signal is interrupted, most of the current methods are to adopt fully connected neural network, but this can only deal with the mapping relationship between input and output at a single moment with being ignorant of the dependence of errors on the past values of strapdown inertial navigation system (SINS), a method of using recurrent neural network (RNN) to assist the integrated navigation system is proposed. RNN can train on the basis of the current and past position and speed samples respectively, so that the neural network can better process the timing signals in the system, and predict the position and speed errors of SINS more accurately. UAV flight test data are adopted to verify the method. The result shows that the navigation accuracy of this algorithm is increased by 77% on average when the satellite signal is interrupted, and this meets the realtime requirements required by navigation. Compared with the traditional RBF neural network assisted integrated navigation system, the root mean square error of its position and velocity is reduced by 39% on average.