A clock synchronization/calibration system combining EKF and LSTM neural networks

Research on methods to improve the timing accuracy and timekeeping capability of timing signals under satellite timing. By using a crystal oscillator counter, the crystal frequency information at the moment of each second pulse is recorded; the recorded historical information is input into an extended Kalman filter (EKF) for filtering, to eliminate the random error of the satellite second pulse signal, and extract the accumulated time of the first $N$ seconds of the BeiDou satellite pulse ${t_{CN}}$, and the crystal oscillator frequency $fre(k)$ at time $k$; and the crystal oscillator change rate $v(k)$ at time $k$; and the historical data output by the EKF is used as the training set, input into the long short-term memory (LSTM) network to establish a prediction model; the algorithm parameters are debugged using the control variable method to find the most suitable prediction model. The experimental results show that the maximum error of the timing signal output by the timing algorithm is 34 ns; the cumulative error of the timing algorithm in 8 hours is 1.001 μs, and the average error is less than 0.125 μs/h. This effectively improves the timing and timekeeping accuracy of the system.