On the economical point of view, cold rolling mill process is one of the most important processes in the integrated steel works. In this thesis, the strip shape from cold rolling mill is predicted with neural network to allow better control of shape parameters. Because of its nonlinear behavior, the strip shape control system in the rolling process does not lend itself to analytic modeling.
Several architecture of the neural networks were studied through a systemetic analysis of acquired data to overcome the inherent constraints in the data collection. The design was also influenced by the need for online adaptation required for real application.
We have investigated intensively two types of network - Multi Layer Perceptron(MLP) and Gaussian Basis Function(FFNN), Time-window FFNN, Time Delayed Neural Network(TDNN), and Gaussian-window TDNN models which are known to be capable to predict time series. GaBF, on the other hand, was tried with Resource Allocating Network(RAN). GaBF is characterized by its neurons each of which is sensitive to a specific region in the input space.
Judging from the experiments, concluded is that RAN is more adequate to strip shape prediction than MLP architectures in terms of prediction quality and learning speed. A strip shape emulator based on the above result was designed and tested to show and acceptible result.
