Prediction of tandem cold-rolled strip flatness based on Attention-LSTM model

Abstract

Flatness is a key quality indicator of tandem cold-rolled strip. Tandem cold-rolled production is a multi-stand simultaneous rolling process. The prediction of flatness is a typical spatial sequence data prediction problem by considering not only the complex generation mechanism but also the spatial dimension dependence. However, the currently available mechanistic and machine learning models for flatness prediction only focus on its highly nonlinear characteristics and ignore its complex spatial correlation. In addition, the over-parameterized “black box” nature of machine learning models has often been questioned. Based on this, a long-short-term memory model with an attention mechanism (Attention-LSTM) is proposed in this paper. The model is structured as a two-layer LSTM network to fully learn the highly nonlinear and complex spatial data correlation of tandem cold-rolled strip flatness. Furthermore, attention mechanisms are also added to the spatial dimension and flatness feature vector dimension, respectively, to enhance the interpretability of the model. The superiority of the proposed model is verified by comparing the error back propagation neural network model (BPNN), the extreme gradient boosting algorithm model (XGBoost), the deep neural network model (DNN), the LSTM model without the attention mechanism, and the Attention-LSTM model with the same dataset and the same optimum training method. The interpretation of the attention mechanism weights is consistent with the mechanistic model, which enhances the credibility of the model. The results of the ablative experiments also validate the effectiveness of incorporating the attention mechanism in the model.