Prediction Model For Longitudinal Surface Crack Of Wide Slab Based On Artificial Neural Network

Longitudinal surface crack is one of the common surface quality defects of continuous casting slabs. The slight longitudinal surface cracks have less influence on the next processing by final finishing, but the severe longitudinal surface cracks would cause serious quality problem and reduce metal yield, even leading to the production emergency. Therefore, in the continuous casting, it is greatly significant to make on-line prediction on the slab quality for ensuring the continuity of production, improving product quality and lowering production costs.Continuous casting is the important process to ensure the steel quality, and manual control becomes more difficult due to the complicated and various process parameters. The Artificial Neural Network possesses strong handling capacity of nonlinear problems and error-tolerance performance, which has been paid more attention because the real-time applications and online response can be achieved.This paper has developed a prediction model for longitudinal surface crack of wide slab based on artificial neural network technology, and the predicted results has been verified and analyzed. Furthermore, an SQL Server database for surface quality of peritectic steel has been also established in order to record the practical data in Najing Steel Company. The following conclusions have been made.(1) The main factors affecting the longitudinal surface crack are the composition of molten steel ([C], [Mn], [P], [S], [Si]), overheat temperature, casting speed, flowrate of cooling water, width dimension of slab, immersed depth of nozzle and water rate for secondary cooling zone.(2) A prediction model for longitudinal surface crack of wide slab based on artificial neural network technology has been developed, with the network parameters of 12 hidden nodes,0.4 learning efficiency, and 0.9 inertia coefficient. The error ratio is less than 10%, and the hit rate is about 90%.(3) The established SQL Server database for quality surface of peritectic steel can realize the collection, storage, update and searching of industrial data, and the preparation for surface quality prediction of subsequent slabs can be made. (4) The selection of sampling data is significant to the model accuracy. The model accuracy can be improved by increasing the data species, ensuring continuity and uniformity of collecting data.(5) The actuality of model and adaptability to industrial production can be ensured by proceeding off-line training of neural network on the basis of replacement of sampling data.