Abstract
Introduction. Accurate state and electromechanical actuator fault estimation is essential for ensuring reliable navigation and fault-tolerant operation of unmanned aerial vehicles (UAVs) operating under noisy and uncertain environments. Problem. Conventional extended Kalman filter (EKF)-based approaches provide efficient recursive estimation for nonlinear UAV systems; however, their performance may deteriorate in the presence of strong nonlinearities, measurement noise, and electromechanical actuator degradations. Goal. To develop a hybrid bidirectional long short-term memory extended Kalman filter (Bi-LSTM-EKF) framework for joint UAV state and electromechanical actuator fault estimation using noisy global positioning system and inertial measurement unit measurements. Methodology. A nonlinear quaternion-based UAV model augmented with electromechanical actuator fault dynamics is formulated and an augmented EKF is designed for joint state and fault estimation. To improve estimation accuracy, a Bi-LSTM network is trained offline using EKF innovation sequences and integrated into the estimation framework to generate adaptive corrections to the EKF estimates. Results. Numerical simulations under progressive and abrupt electromechanical actuator fault scenarios demonstrated the effectiveness of the proposed Bi-LSTM-EKF framework. Compared with EKF and unscented Kalman filter approaches, the proposed method reduced the position root mean square error (RMSE) by approximately 46–48 % and achieved lower fault estimation errors, with fault RMSE values of 0.0765 and 0.0826 for the progressive and abrupt fault scenarios, respectively. These results confirm the improved trajectory reconstruction and actuator fault estimation capability of the proposed framework under noisy operating conditions. Scientific novelty. The integration of Bi-LSTM-based innovation learning with an augmented EKF for the simultaneous estimation of UAV states and electromechanical actuator faults. By combining model-based recursive estimation with bidirectional temporal learning, the proposed framework captures nonlinear fault evolution and dynamic transitions more effectively than conventional filtering techniques. Practical value. The proposed framework provides an effective solution for UAV navigation, electromechanical actuator health monitoring, fault diagnosis, and fault-tolerant flight control applications requiring reliable state and fault estimation under uncertain operating conditions. References 29, tables 2, figures 7.