Degree Name

Doctor of Philosophy


Department of Mechanical Engineering


The quality control in hot strip mills usually include thickness, width, shape, flatness, yield strength, tensile strength, ductility, etc.. This research takes the width control at the roughing mill of a hot strip mill as an example of AI applications for the quality control of the hot strip. The width control consists of the bar-to-bar width control for the mean width of bars, the in-bar width control for the width variation within bars and the head/tail width control for reducing the crop loss at the crop-shear in front of the finishing mills.

Back-Propagation (BP) neural networks and Radial Basis Function (RBF) neural networks were applied to model and predict the bar-to-bar width and the in-bar width. The training and the testing of the neural networks are based on the data collected from the BHP Steel SPPD roughing mill. Bias and momentum are used in the learning process of BP networks for speeding up the learning process. Also Cauchy method is applied here to avoid the local minimum of the learning process. Moreover, polynomial regression models were developed for the comparison between the neural network models and the regression models. The comparison is also conducted for the performances between the BP networks and the RBF networks. The factor analysis both by BP networks and the regression models are used to show the most effective parameters for the width variation.

The performance of the head & tail stroking system is dependent on the original head/width shape of the incoming bar and the trace of the edger roll for the stroking. The original shapes can be classified as a limited number of basic shapes. On the other hand, the trace of the edger roll for the stroking can be represented by three parameters. Based on expert knowledge, BP neural networks are set up for representing the relationships between the basic shapes and the stroking parameters. Well trained networks can be used to automatically identify the original shapes and then provide the suitable parameters for the stroking system. The results of training and testing by some data from the BHP Steel roughing mill are provided. General Regression neural networks are also proposed here for the comparison between these two kinds of networks.

An approach of Fuzzy Logic Controllers (FLC) is suggested for the bar-to-bar width control. In the set-up (learning) process of the FLC , it is found that BP method can lead to the failure of the learning. In order to overcome this problem, a hybrid learning algorithm is proposed for the set-up of the FLC. This algorithm includes a self-organised learning phase, a supervised learning phase and a mixed learning phase. The methods of the self-organised learning used here are the competitive learning and the input-output product-space clustering. This hybrid methodology is used to generate initial fuzzy rules and initial membership functions, to adjust membership functions and to refine final fuzzy rules. It is applied to the learning of the FLC for the bar-to-bar width model.

Guide-lines to develop the expert system for all quality items at the hot strip mills are provided in this thesis. Under these guide-lines a prototype expert system is developed for the width control of the roughing mill of a hot strip mill. This expert system is built on G2 system introduced by Gensym Corporation. Neural network models for the bar-to- bar width, in-bar width and head/tail width can be deployed to C programs which are recalled by the expert system. A rule-based system for the roughing mill width monitoring is integrated into this expert system. The results of the test running of the expert system by the raw data collected from the roughing mill are shown through the photos of the main window work-spaces provided in this thesis.

Further study on applications of BP neural networks is included in this thesis. It addresses the choice of architectures, the choice of training strategies and the problems of convergence and generalisation. The results of the study confirm that the outcome of the learning for BP neural networks is related to many factors. These factors include the formations of input and output variables for BP networks, the number of layers and nodes per layer, the type of transfer functions, the learning rule, the sample sizes of training and testing, as well as the number of training cycles. The guidelines are developed to obtain good results for the applications of BP neural networks.