System Identification Based on an Improved Generalized ADALINE Neural Network

This paper presents an online system identification method for a linear time-varying system whose parameters change with time. The method is based on an improved generalized ADAptive LINear Element (ADALINE) neural network. It is well known ADALINE is slow in convergence which is not appropriate for online application and identification of time varying system. To speed up convergence of learning and thus increase the capability of tracking time varying system parameters, two techniques were proposed, i.e. i) a momentum term added to the weight adjustment and ii) training on a sliding window over data set. While the momentum speeds up convergence, it also shows over-shooting and while the sliding window training helps to track variable parameters better but also tracks noise closely. An average weight adjustment and dual epoch learning are proposed to improve performance. Simulation results show that the proposed method provides indeed faster convergence and better tracking of time varying parameters.