Research on dynamic rolling force prediction model based on CNN-BN-LSTM

Abstract

Considering the effects of vertical vibration displacement and horizontal vibration displacement of work rolls, a CNN-BN-LSTM dynamic rolling force prediction model based on the combination of one-dimensional convolutional neural network (1D-CNN), Batch Normalization (BN) and Long Short-Term Memory Network (LSTM) is proposed, and the model is optimized by Adam optimization algorithm for different parameters by designing independent adaptive learning rates. The model firstly uses CNN to extract data features effectively, secondly uses BN algorithm to improve the training speed, suppress gradient disappearance and reduce overfitting, and finally uses LSTM to extract data signals of time series to construct the dynamic rolling force prediction model. The proposed model is calculated by collecting PDA (process data acquisition) production data, vibration signals of horizontal and vertical vibrations of work rolls from the second finisher (F2) of a hot tandem rolling mill. The results show that the prediction accuracy of CNN-BN-LSTM is 90.68% and is higher than that of the Pei-Hua Hu model and CNN model, which verifies that the CNN-BN-LSTM model has strong generalization ability, fast training speed and high prediction accuracy. The relationship between process parameters and vibration is quantitatively analyzed based on the CNN-BN-LSTM model. The results show that reducing the rolling speed, reducing the entrance thickness and increasing the exit thickness can improve the stability of the rolling process.