Fusion of feature extraction network and multi-head attention based instability state classification in continuous casting

Abstract

Continuous casting is a core process in modern steel production. However, due to the complex coupling of multiple factors, various instability states, including cold-tooth, excessive hydrogen content, and high sulfur content, are frequently observed. Such instabilities may lead to steel breakout accidents in severe cases, compromising both product quality and production safety. Therefore, instability state classification is required to guide operators in implementing targeted preventive measures and optimize the production process. Traditional instability state classification methods rely on expert experience or machine learning models, and their robustness and feature extraction capability are limited under noise interference. To address these limitations, a new instability state classification method is proposed for continuous casting, which integrates a feature extraction network with multi-head attention mechanism. Features related to instability states from temporal data are extracted using the gated recurrent unit network, while feature dependencies are captured in parallel from multiple subspaces by the multi-head attention mechanism. This fusion enhances the model focus on instability-related features and improves its noise resistance, enabling high-accuracy and robust classification of instability states in continuous casting. Experiments conducted on real datasets from a multi-mode continuous casting and rolling production line demonstrate that the proposed method performs excellently across various instability state classification tasks.