基于卷积神经网络的热轧薄板力学性能建模

Modeling of Mechanical Properties of Hot Rolled Sheet Based on Convolutional Neural Network

目的为了提高热轧薄板力学性能的预测精度,采用大数据与卷积神经网络相结合的方式建立高精度的预测模型。方法建模前,对工业大数据进行预处理,包括去除异常值、聚类、均衡数据以及归一化,以得到高质量的数据集。同时,采用贡献权重法对输入参数进行筛选,去除弱相关的变量以降低模型的复杂程度。在此基础上,采用LeNet-5结构建立卷积神经网络并优化模型的超参数。结果最终建立了热轧薄板力学性能预测模型,该模型对屈服强度的预测误差基本保持在−7%~8.5%,对抗拉强度的预测误差基本保持在−5%~6%,表现出较高的预测精度。结论将卷积神经网络模型与传统的BP神经网络模型进行了预测对比,发现卷积神经网络能够利用其局部连接的优势给出更高的预测精度。

The work aims to combine big data with convolutional neural network to establish a high-precision prediction model to improve the prediction accuracy of mechanical properties of hot rolled sheet.Before modeling,industrial big data was preprocessed,including outlier removal,clustering,data equalization and normalization,to obtain high-quality data sets.At the same time,the contribution weight method was used to filter the input parameters and remove the weakly correlated variables to reduce the complexity of the model.On this basis,the convolutional neural network was established with LeNet-5 structure and the hyperparameters of the model were optimized.At last,a prediction model for mechanical properties of hot rolled sheet was finally established.The prediction error of the model for yield strength was maintained between−7%~8.5%,and the prediction error of tensile strength was maintained between−5%~6%.The prediction results showed high prediction accuracy.In addition,the comparison of the results of convolutional neural network and traditional back propagation neural network finds that the convolutional neural network can give higher prediction accuracy by taking advantage of its local connection.