Strip deviation analysis and prediction based on time series methods in hot rolling process

Abstract

Strip deviation presents a significant challenge in hot rolling processes, affecting both product quality and manufacturing efficiency. Currently, most of the strip deviation correction operations rely on manual adjustments, which are labor-intensive and error-prone. This study pioneers the integration of a strip deviation measurement system with a time series prediction model to predict strip deviation and provide operators with timely warning signals. It introduces a novel time series prediction model utilizing dual attention mechanisms: one to identify feature-level correlations and another to capture temporal-level dependencies and patterns. An optimized version of the traditional Multi-Head Attention mechanism, named Compact Multi-Head Attention, is incorporated. To further boost the model's predictive accuracy, a shuffle operation is also integrated. Additionally, the dataset is augmented with rolling force difference and roller gap difference, based on an analysis of strip deviation principles, leading to notable improvements in predictive accuracy. Comprehensive testing with actual data from a hot strip mill confirms the model's outstanding performance in predicting strip deviation, surpassing several baseline models. The results highlight the effectiveness of this approach in strip deviation prediction in industrial environments.