Data-Driven Soft Sensor for Hot Strip Mill Process Based on Multiscale Information-Fusion Attention Network

Mechanical properties (MPs) of the hot strip mill process (HSMP), as key indicators of product quality, are always difficult to measure online due to cost or technological limitations. The development of a soft sensor is a viable solution that provides a steady and reliable online estimation of MPs based on available process variables. In this article, a data-driven soft sensor approach, named multiscale information-fusion attention network (MIAN), is proposed for MP prediction of the steel strip in HSMP. Specifically, an intrinsic property-guided feature extractor (IPFE) is developed, which groups variables according to their intrinsic properties, extracts corresponding features respectively, and integrates them into global shared features. This will effectively avoid information confusion caused by directly extracting features from mixed variables. In addition, we propose a novel attention-based MP-specific weight generator (AMWG) to yield MP-specific weights to emphasize variables that contribute to specific MP predictions, which will be further used to generate local attention-weighted MP-specific features. With the joint of the global shared features and the local attention-weighted MP-specific features, the multiscale fusion features are obtained and are used as the input of the newly designed multiscale information fusion-based predictor (MIFP). When combined with the proposed two-stage training strategy, MIAN can be learned and produce superior MP prediction performance. Extensive experimental results on HSMP data collected from a real Iron and Steel Company Ltd., China, elucidate that the proposed soft sensor outperforms the state-of-the-art methods in terms of root mean square error (RMSE), mean absolute error (MAE), and R-square ( $R^{2}$ ).