Abstract
We present a new identification algorithm for parameter estimation of time-varying linear discrete-time systems. The change of environment of the system might cause system parameter fluctuations. As these fluctuations are unknown, the system parameters can not be represented by a specific function. The artificial neural networks which have the ability to learn complex nonlinear relationships, are applied to represent the fluctuations of system parameters.
These system parameters are estimated by optimizing the nonlinear objective function which is the total error of the system and the model. The optimization problem of the objective function results in the learning of the interconnection weights of the network. Some techniques are required for setting appropriate initial points in optimization, because there is many unfeasible local minimum points. To find the globally optimal set of weights, the weights that result from training the more simplified function are used to initialize the networks.
Furthermore, it is verified by simulation that the effectiveness of the proposed method can be applied to the system with unknown parameters.
