A Method of Rapid Training for Neural Networks Based on Kalman Filter

According to the requirement for real-time modeling in industrial processes, a rapid and efficient method based on Kalman filter (KF) for training neural networks (NN) was presented. In this algorithm, the weights of hidden-layer were initialized randomly at the beginning of training and left unchanged, while the weights of out-layer were served as the states of an ordinary Kalman filter and adjusted automatically according to real-time input-output data of dynamic systems. It considered NN training as the problem of linear state estimation. Simulation results for a non-linear multi-input and single-output (MISO) system showed that the proposed training algorithm was rapider and more efficient than back-propagation (BP) and extended Kalman filtering (EKF) algorithms. Therefore, the proposed algorithm is more suitable for on-line learning compared with BP and EKF