5th International Conference on Advances in Solidification Processes (ICASP-5) 5th International Symposium on Cutting Edge of Computer Simulation of Solidification, Casting and Refining (CSSCR-5)

A method for detecting sticker breakouts in continuous casters is proposed based on extensive feature engineering and advanced analytical modelling. It is observed that the temperature patterns generated during a sticker breakout have a consistent signature, owing to the underlying physical processes involved. The aim of the present exercise is to capture this latent physical phenomenon in a reliable and robust way. The temperatures measured by thermocouples in the mold are used to extract features of physical significance and operational importance. The features were custom built iteratively to capture the difference between true and false patterns. These features are consequently used to develop a Gradient Boosting model to detect sticker breakout. The model is trained on previously raised sticker alarms that are manually tagged as true or false alarms for a time period of around $3$ years ($1500$ alarm files). The sample space of the non-alarm (tagged $0$) is increased by using data collected during normal operation. The GBM model shows an overall superior performance as compared to the existing in-place logic. While ensuring that no true alarms are missed, the GBM model reduces the false alarms by around $80$%. The reduction in false alarms imply a huge production advantage while the feature engineering involved in the modelling process makes this a unique exercise in the realm of continuous casting research. The data used in the analysis was generated in continuous casters operating in Tata Steel Jamshedpur.