Time‐varying parameters identification for dual‐control aircraft based on efficiency learnable extended Kalman filter

Abstract An efficient identification method is proposed to estimate the time‐varying parameters of a nonlinear dual‐control aircraft. Based on the Extended Kalman Filter (EKF), the long‐short‐term memory (LSTM) neural network is introduced to adjust the filter gain to reduce the filter error caused by model uncertainty and linearization. To solve the heavy training burden of the stochastic gradient descent (SGD), RMSprop, and Adam training algorithm, an efficient training algorithm is presented, which uses Independent Extended Kalman Filter (IEKF) to transform the network weights training problem into an optimal weights identification problem to achieve the convergence speed of second‐order training. To further reduce the training burden of the network, the network weights Matrix Factorization (MF) is introduced to reduce the number of learnable weights. In addition, the training performance of the LSTM neural network using MF structure and IEKF training algorithm is analyzed by the Lyapunov method, and the selection method of the artificial process noise matrices is given. Time‐dependence data fitting experiments show that the efficient LSTM neural network proposed herein can greatly reduce the training burden of the network, and simulation results show that the Learnable Extended Kalman Filter (L‐EKF) achieves better performance than some existing filtering methods in the time‐varying parameter identification of the dual‐control aircraft.