Data-Driven Soft Sensor Based on Sparse Relational Graph Attention Network for Hot Strip Mill Process

Abstract

Due to technological or financial constraints, it is always challenging to measure the mechanical properties (MPs) of the hot strip mill process (HSMP) online. A feasible solution that offers a consistent and trustworthy online estimation of MPs based on available process variables is to develop an Effective soft sensor. To achieve this, we propose a novel method for MP modeling based on a sparse relational graph attention network (SRGAT). In SRGAT, a sparse relationship learning (SRL) module is first utilized to learn a graph structure that can describe sparse interactive correlations among process variables. Then, both the learned graph structure and process variables are fed into the graph attention network (GAT) to highlight useful information for MP prediction of the steel strip. The proposed SRGAT soft sensor surpasses state-of-the-art approaches in terms of root mean square error (RMSE), mean absolute error (MAE), and r-square (R2), according to experimental results on the HSMP data collected from a real Iron & Steel Co., Ltd, China.