Original Article

Application of deep learning and attention mechanisms in continuous casting breakout prediction

¹ School of Mechanical Engineering, Yancheng Institute of Technology, Yancheng, PR China
² Material Technology Research Institute, Hebei Iron & Steel Group, Shijiazhuang, PR China
³ School of Transportation, Inner Mongolia University, Huhehaote, PR China

^* e-mail: zhangbenguo@ycit.edu.cn

Received: 1 July 2024
Accepted: 5 March 2025

Abstract

To improve the accuracy of a steel breakout prediction system, a steel breakout prediction model based on Bayesian optimisation (BO) of convolutional neural network (CNN)-bidirectional gated recurrent unit network (BIGRU) and multi-head self-attention mechanism (MA) was proposed for the thermocouple temperature characteristics. The essence of the thermocouple temperature measurement method was analysed by examining the single-couple time-series characteristics and group-couple spatial linkage characteristics of thermocouples. The essence was the problem of pattern recognition of dynamic temperature characteristic waveforms. CNN was used to extract spatial characteristics of the data and BIGRU was employed to extract time-series characteristics to construct the CNN-BIGRU network. Moreover, BO was used to find the optimal hyperparameter combinations for the CNN-BIGRU network to determine the optimal hyperparameter combinations for the CNN–BIGRU network and MA was introduced to improve prediction accuracy. A network model based on deep learning and attention mechanisms was finally developed and applied to the field of continuous casting breakout prediction system. The breakout prediction model was tested in conjunction with actual continuous casting production data. The results show that the accuracy of this breakout prediction system is 99.5% and the reporting rate is 100%.