Modeling of breakout prediction approach integrating feature dimension reduction with K-means clustering for slab continuous casting

Abstract

Breakout is a catastrophic accident in continuous casting. The existing breakout prediction methods based on logical judgment and neural networks need to constantly adjust the prediction parameters or prepare high-quality samples as the input, resulting in poor robustness and unstable precision stability. Therefore, it is particularly important to develop a breakout prediction method that not only can predict breakout accurately but also avoid human intervention significantly. This work proposes a novel approach for breakout prediction combining K-means clustering and feature dimension reduction. The method uses feature dimension reduction to obtain the typical feature vector (TFV) that can characterize the original temperature change trend, and then a K-means clustering model is established to realize online detection of breakout prediction. The results show that the model has a 100% alarm rate for the true breakout, and meanwhile, reduces the number of false alarms from 555 to 217 compared with the on-line breakout prediction system (BPS). The proposed method does not need to adjust the prediction parameters frequently or prepare the input samples carefully, which not only avoids the human intervention but also meets the requirements of online monitoring for the practicality and applicability of the breakout prediction method.