A dynamically generated fuzzy neural network and its application to torsional vibration control of tandem cold rolling mill spindles

Based on the model of Higgins and Goodman, we describe a dynamically generated fuzzy neural network (DGFNN) approach to control, from input–output data, using on-line learning. The DGFNN is complete with the following powerful features drawn or modified from the existing literature: (1) a small FNN is created from scratch—there is no need to specify initial network architecture, initial membership functions, or initial weights, (2) fuzzy rules are constantly combined and pruned to minimize the size of the network while maintaining accuracy, irrelevant inputs are detected and deleted; and (3) membership functions and network weights are trained with a backpropagation-type algorithm. We apply the DGFNN controller to a real-world application of controlling the torsional vibration of tandem cold-rolling mill spindles with a simulated plant. The results of the DGFNN controller are compared with the performances of a conventional proportional-integral controller and a neural controller using recurrent cascade correlation with quickpropagation through time. We show that while both neural approaches increase the control precision and robustness, the DGFNN controller gives the best results for reducing the speed deviation and suppressing the torsional vibration of the spindles, as well as is more computationally efficient.