ECCC 62用于小方坯质量控制的神经元软件.pdfVIP

Quality control and detection of defects Session 19 1 A Neural Network Model fo r Billet Quality Control 1 2 1 1 1 Stefan Räfså , Jukka Laine , Matti Tukiainen , Suvi Rannantie , Päivi Tikkanen 1) FNsteel Oy Ab, Koverharintie 303, FI-10820 Lappohja, Finland 2) Aalto University, School of Chemical Technology, Vuorimiehentie 2, FI-02015 TKK, Finland Stefan Räfså, FNsteel Oy Ab, Koverharintie 303, FI-10820 Lappohja, Finland, phone +358 20709 4176, Fax +358 207094150, stefan.rafsa@ Key Words: Billet, Quality control, Continuous casting, Neural networks, Bayesian methods Billet quality control is a complex process dealing with many continuous casting process parameters, criteria and exceptions. Moreover, today customers require increasingly shorter delivery time. All delays in steel production chain have to be eliminated. Koverhar’s quality control system is continuously being developed in order to improve its speed, accuracy and reliability. This work presents a Bayesian MLP neural network model, called mould model, which is a part of the quality system. The mould model was built by training a neural network with the empirical data set based on customer feedback. The mould model quickly tracks any mould level disturbances that may reduce billet surface quality and thus increase the risk of billet downgrading. As input the mould model uses different features calculated from two online process signals at each casting strand: mould level and a stopper rod position. As an output the model gives a quality index for every billet produced. Validation of the model confirms that the model gives accurate and reliable predictions. As a result it helps the quality control function to operate in an ob