Application of variational autoencoder network to real-time prediction of steel crown in the hot strip rolling mill process

In the hot strip rolling mill (HSRM) process, accurate prediction and control of the strip crown are critical for quality assurance. In order to cope with this challenge, this study designed a real-time prediction and update system of strip crown based on the cloud-edge-end collaboration framework. Firstly, this work optimizes the traditional variational autoencoder (VAE) network by refining the loss function structure to improve feature extraction and prediction, tailoring the VAE to the unique requirements of crown prediction. Second, according to the characteristics of multi-stand distribution in the HSRM process, a distributed framework is constructed to enable distributed extraction and fusion of crown-related features, generating predictions based on the fused features. Additionally, to adapt to different strip specifications, a global and local update method is proposed to dynamically optimize model parameters, marking a notable advancement in adaptability for real-time industrial applications. The application results from two actual HSRM production lines (2150 and 1580 mm) demonstrate that the proposed method can decrease the prediction error to 2.650 μm on average. Finally, by using a cloud-edge-end prototype system with a 50 ms sampling interval, the system enables real-time prediction and supports online local model updates, significantly improving traditional methods while enhancing both operational efficiency and quality control.