Utilization Research Of Neural Networks For The Caster Surface Quality Prediction Model

With the development of continuous casting technique, the require for the variety and quality of the steel products becomes more and more strict, how to improve the quality of the continuous casting slab has become an important problem of the continuous casting technology research. Most steelworks introduce the computer into the continuous casting produce process, and have acquired huge achievement. The casting slab surface quality prediction expert system is one important portion in this field.In the actual product process of Meishan Steel works, the surface longitudinal crack takes a big part in the disfigurement casting slabs, results in descending the ratio of the finished product. To solve this real produce question, the craft parameters in the real product process and other factors that causes surface longitudinal crack in the slab is considered. After analyzing the craft data from the real product process of Meishan Steel works, the casting slab surface quality prediction model is established that can predict the probability of the surface longitudinal crack in the casting slab. This model adopts the Back-propagation algorithm as the mathematics theory of this system, and develops by the Visual Basic 6.0.This model can be used to reduce the surface longitudinal crack through adjusting the correspondence craft data. All of the factors which influent the slab quality of this model are considered in these training data. It covers the factors so widely that the ratio of nicety is higher.The causes of surface longitudinal crack in the slab and the factors which influent the slab quality are all analyzed, and the study training data are constituted by all the factors which influent the slab quality in this model. After analyzing and filtrating the data from the data collecting system, the structure of training data are confirmed. With the consideration of the character of the Back-propagation neural networks with the practice in real product process, the structure and performance parameters of the Back-propagation neural networks is confirmed, and the neural network prediction model is established. Then the study training data should be input into the prediction model to study this model, at last this model will give the judgement result of the possibility of surface quality question by means of metallurgical theories and expert experience. This model has been verified off-line using historical data, and the results indicated this model could exactly predict the surface longitudinal crack of the HP295 up 90%.The standard Back-propagation algorithm is widely applied in the fields of the artificial neural networks. To change the study rate and add momentum item to the standard Back-propagation algorithm will be the approved Back-propagation algorithm.The approved algorithm which can improves the learning speed of this model is certificated.After analyzed by this model we find out that the main reason of the surface longitudinal crack of HP295 is the irrational distribution ratio of the secondary cooling water, so the formation of the surface longitudinal crack can be reduced in the production by adjusting the distribution ratio of the secondary cooling water.