采用神经网络预估法建立板材力学性能预测模型

Prediction-Evaluation Model of Mechanical Properties of Steel Strip Based on Artificial Neural Network

为提高产品质量,降低产品成本,开发了板材的屈服强度、抗拉强度、延伸率等力学性能的预测模型;介绍了建立热轧带钢力学性能质量模型的数据挖掘过程;用普通神经网络建立起由工艺参数预测力学性能的质量模型,模型预测结果的5%命中率是0.508;然后,提出一种新的建模方法——预估法,该方法是以三层BP神经网作为基本模型,通过增加模型层数,缩小底层子模型的预测范围,从而提高模型的泛化能力,这种方法的关键问题是能够对测试数据实现正确分类;利用该方法建立起质量模型,预测结果的5%命中率达到0.706,完全可以满足现实生产需要。

In order to improve the product quality and reduce the product cost, the mechanical properties quality models of yield strength, tensile strength, and elongation are established. The data mining process of establishing the quality model, that could predict the mechanical properties of hot--rolled steel strip with the technological parameter, is introduced. Then, the quality model whose hit ratio of 5% deviation reach 0. 508 was established by applying the technology of basic artificial neural network. The technology of prediction--evalua tion is proposed. This method increases the generalization ability of the model through increasing model layers and decreasing the prediction of submodel of base layer basing on the basic model of artificial neural network. The key of the method lies in that whether the tested data can properly enter the submodel or whether the tested data can be properly classified. Finally, the quality model whose hit ratio of 5% deviation reach 0. 706 was established by applying the technology of prediction--evaluation, and this model could meet current industrial demand fully.