Mechanism-Constrained Multistage Recursive Soft Sensor Framework for Slab Temperature Prediction in Reheating Furnace

Abstract

Slab temperatures are difficult to measure directly in real time during the heating process in the reheating furnace. This article analyzes the slab heating process as a multistage manufacturing system (MMS) and proposes a soft sensor framework named mechanism-constrained multistage recursive network (MC-MRN) to predict slab temperature. The proposed method first uses the mechanism model to generate label data for each stage as the pretraining basis and introduces a new prediction regularization term in its loss function using the mechanism information to guide and constrain the feature extraction process so that the extracted features more comprehensively reflect the original data and its stage information. Furthermore, based on the physical relationship between stages, we connect the pretrained models of each stage in series and input the features containing the production information from all stages into the prediction network for fine-tuning, thereby constructing an overall soft sensor model framework. This design ensures that the framework structure aligns with the physical structure of the reheating furnace, and the constraints of the mechanism model enhance the interpretability and reliability of the framework, ensuring that its predictions remain within a reasonable range consistent with the laws of physics. The experimental results show that the MC-MRN soft sensor framework, after fine-tuning, demonstrates high accuracy in slab temperature prediction.