A Flatness Predict Model Based on Deep Belief Network for Steel Rolling Process

Abstract

The flatness deviation of the cold-rolled strip steel sheet reflects the product quality of strip steel, and it is a vital quality index to monitor in the cold rolling strip process. Given the problem of delayed detection of strip rolling at low speed, a practical and effective flatness predict model is established by a data-driven method, which integrates Mutual Information (MI) with Deep Belief Network (DBN) in this thesis. In the process of feature selection, MI technology is applied to the model, which can describe the quantitative relationship between flatness variables and process variables. Furthermore, it is the criterion for selecting the variables as the best input. Then, DBN is presented as the flatness predictor, it can reinforce the ability to extract features from raw data which with the character of high dimension and nonlinear. Firstly, the Contrastive Divergence learning (CD) is applied to pretraining the model by training the Restricted Boltzmann Machine (RBM) which is the basic unit of DBN. Then, a global back-propagation (BP) algorithm is used in the whole model’s fine-tuning, and the whole model parameters were optimized by Adaptive moment estimation (Adam) finally. As a result, the model’s performance, mean square error (MSE) is about matching 0.0009, which is better than the prediction model formed by BPNN or Stacked Auto Encoder (SAE). (Abstract)