Deep learning-based flatness prediction via multivariate industrial data for steel strip during tandem cold rolling

Abstract

Flatness deviations in the tandem cold-rolling process of steel strips have a direct impact on product quality and shape, leading to strip breakage, reduced working speed, and equipment damage. However, conventional physics-based numerical models are inadequate for accurately predicting flatness in the complex operating conditions and variables of tandem rolling environments. To address this challenge, a novel approach is proposed that utilizes deep convolutional neural networks (DCNNs) based on real industrial data from tandem cold rolling. The multi-input and multi-output architecture of our DCNNs enables them to solve the multi-level nonlinear problem associated with flatness prediction in the tandem cold-rolling process. The flatness profiles are effectively predicted using the proposed method, incorporating multiple variables without requiring additional data pre-processing methods. Additionally, the effects of network width, depth, and topology on flatness prediction performance are thoroughly investigated. The developed Inception-ResNet demonstrates remarkable predictive performance while using fewer model parameters and exhibiting lower computational complexity compared to other network architectures. Specifically, the proposed Inception-ResNet-39 model, consisting of 39 layers of learnable parameters, achieves state-of-the-art predictive performance. Our deep learning-based approach accurately predicts flatness in tandem cold-rolling through end-to-end modeling and provides complete pipelines for model transfer construction to ensure efficient implementation.