The condition monitoring of switch machines (SM), which are critical actuators in railway turnout systems, is essential for ensuring the safe and reliable operation of railways. However, traditional Extended Kalman Filters (EKF) face challenges in effectively capturing high-order nonlinear characteristics and accurately modeling the dynamic behaviors of short-term periodic systems. To address these limitations, this paper proposes a Deep Extended Kalman Filter (DEKF) framework designed specifically for short-term periodic nonlinear systems. The proposed approach integrates a Long Short-Term Memory (LSTM) network to extract intermediate features from the EKF, thereby enhancing the utilization of nonlinear dynamic information. Furthermore, a time-series sample set covering complete operation cycles is constructed to improve the representation of prior knowledge embedded in short-term periodic actions of the switch machine. By employing an offline training and online fusion strategy, the DEKF achieves high-precision real-time state estimation of the switch machine system. Simulation experiments conducted on a nonlinear switch machine model demonstrate that the proposed DEKF significantly outperforms the traditional EKF in state estimation accuracy, providing strong theoretical support for real-time condition monitoring and fault diagnosis of railway signaling systems.

Deep neural network,extended Kalman filter,state estimation,switch machine