(燕山大學 工業(yè)計算機控制工程河北省重點實驗室，秦皇島 066004)
摘 要: 在鋁熱軋過程中，軋制力預報精度直接影響著成品的產量和質量。為了提高鋁熱連軋軋制力預報精度，提出一種基于深度學習方法的多層感知器(Multi-layer Perceptron，MLP)軋制力預報模型。模型利用MLP的函數(shù)逼近能力來回歸軋制力。模型以小批量訓練為基礎，利用Batch Normalization方法穩(wěn)定網絡前向傳播的輸出分布，并使用Adam隨機優(yōu)化算法來完善梯度更新，以解決MLP模型難以訓練的問題。仿真結果表明：模型使網絡預測與實測數(shù)據的相對誤差降低到3%以內，實現(xiàn)了軋制力的高精度預測。
關鍵字: 鋁熱軋；軋制力預測；深度學習；多層神經網絡；優(yōu)化算法
（Key Lab of Industrial Computer Control Engineering Department of Yanshan University, Qinhuangdao 066004, China）
Abstract:In the aluminum hot rolling, the prediction accuracy of the rolling force directly affects the output and quality of the finished product. In view of the inherent defects of traditional rolling force model, a MLP rolling force prediction model based on deep learning method was proposed. The model uses MLP’s function approximation ability to regress the rolling force. Based on the Mini-batch training, the model uses Batch Normalization method to stabilize the output distribution of the network forward propagation, and uses the Adam stochastic optimization algorithm to improve the gradient updating so as to solve the difficult training problem of the MLP model. The simulation results show that the model can reduce the relative error between the network prediction and the measured data to less than 3%. Compared with the traditional mathematical model, this method realizes the high precision prediction of the rolling force, and realizes a high-precision prediction of rolling force.
Key words: aluminum hot rolling; rolling force prediction; deep learning; multilayer neural network; optimization algori
