Flatness prediction of hot-rolled dual-phase steel after laminar cooling using a mechanism-data fusion-driven decomposed attention linear network

Hot-rolled dual-phase steel often suffers from flatness defects after laminar cooling, but high-precision online prediction remains difficult because production data are fragmented across industrial control levels, key internal thermo-metallurgical states are difficult to measure directly, and process variables are strongly coupled and nonlinear. To address this issue, a mechanism-data fusion framework is developed on a multi-source Industrial Internet of Things platform. A coupled heat-transfer and phase-transformation model is first built to generate mechanism-derived virtual features. Then, a deep-learning model termed the Decomposed Attention Linear Network (DALIN) is proposed by introducing multi-head self-attention into the trend and remainder branches of DLinear, while Differential Evolution is used for global hyperparameter optimization. Results show that mechanism-data fusion reduces the average mean absolute error (MAE) and root mean square error (RMSE) by 0.243 and 0.358, respectively, and improves the average coefficient of determination (R 2 ) by 0.038. Under the full framework, the average R 2 reaches 0.918, with an average prediction interval coverage probability (PICP) of 0.947 and an average mean prediction interval width (MPIW) of 2.359. The deployed model achieves 25 ms inference per sample and supports online incremental retraining. Industrial application further indicates that the framework can provide practical support for feed-forward flatness control, with favorable changes observed in flatness-related production indicators during the application period.