Application of neural networks for predicting hot-rolled strip crown

Abstract

In this paper, a prediction model for hot-rolled strip crown based on an artificial neural network (ANN) is presented. Considering the mean squared error (MSE) and correlation coefficient (R), production data of 10,133 coils collected from a hot rolling plant are used to establish models. A feed-forward neural network with one hidden layer, a non-dominated sorting genetic algorithm II (NSGA II) optimized ANN, and a deep neural network (DNN) are applied to evaluate the prediction performance. Parameter settings of the ANN and NSGA II including learning rate, hidden neurons, activation function and population size, and crossover probability, are investigated to acquire the optimal model. The structure of the DNN, including how many layers and units the network should contain, is also studied. Prediction performance comparisons of the ANN, NSGA II-ANN, and DNN are presented. Among the ANN, NSGA II-ANN and DNN, the results show that the DNN has the highest prediction accuracy. Root mean squared error (RMSE) of the proposed DNN is 2. 06μm, and 97.04% of the prediction data have an absolute error of less than 5μm. Through model response surfaces, effects of four key operating parameters are investigated. The results indicate that the proposed DNN with strong learning ability and generalization performance can be well applied to hot rolling production.