Group publication title:
Subject and Keywords:
Principles of employing feedforward artificial neural networks for fast and robust estimation of dynamic system parameters are reviewed briefly. In this approach, the neural network approximates the mapping from the system observation space into the parameter space. It is pointed out that for the conventional neural network architectures the network size and time needed for its training increase quickly with the estimated parameter range and with the approximation accuracy level required. ; Moreover, due to the local minima effect, the training process is likely to become prematurely terminated. To overcome these difficulties, a modular neural network architecture is proposed which comprises classifier and approximator modules, both driven by the system under test observations. With this architecture, the domain of the mapping is partitioned into a number of nonoverlapping regions. ; The classifier makes a decision as to which predefined region the given observation vector belongs to. This information is then used to select an appropriate weight vector (and possibly the structure) of the approximator module, so as to minimise the parameter estimation error locally, within the region identified. A numerical example is presented to show that the proposed approach offers higher estimation accuracy and huge savings in time required for the training.