【作者】 胡于华；

【导师】 程晓茹；

【作者基本信息】 武汉科技大学 ， 材料加工工程， 2006， 硕士

【摘要】 人工神经网络模型具有很强的容错性、自适应性和非线性的映射能力,特别适于解决因果关系复杂的非确定性推理、判断、识别和分类等问题。人工神经网络从实验数据中通过自学习自动获取数学模型方面具有独特的优越性,它无需人们预先给定公式的形式,而是以实验数据为基础,经过有限次的迭代计算,就可获得一个反映实验数据内在规律的数学模型,神经网络尤其适用于处理规律不明确、成分工艺变量多的问题,因此,可以利用人工神经网络实现板带化学成分、轧制工艺参数与力学性能的直接映射,进而达到较好的对板带力学性能的预报。本文通过对神经网络模型的理论依据和建模方法的研究,在利用涟钢CSP热轧Q345B钢种的化学成分和工艺参数与力学性能检验数据的基础上,建立了BP神经网络力学性能预报模型。此模型是一个典型的三层网络结构,即输入层、隐层和输出层。输入层有11个神经元,分别为碳、硅、锰、磷、硫、铝、钙含量及F1机架入口速度、终轧温度、卷取温度和成品厚度,输出层有3个神经元,分别为屈服强度、抗拉强度和延伸率。神经网络模型经过训练,得到的预报结果与实际测量的数据相比较表明:BP神经网络预报产品的力学性能精度较高,该方法具有良好的推广价值。同时,利用训练好的神经网络预报模型,可以分析化学成分和工艺参数与产品力学性能之间的关系,便于发现规律,为产品力学性能的在线预测和产品的轧制工艺优化与调整等有重要的指导意义和参考价值。更多还原

【Abstract】 The Artificial neural network (ANN) model possesses with good fault tolerance, self- adapted and non-linear mapping, especially it is suitable to resolve complex causal relation non-question and so on determinism inference, judgment, recognition and classification. The ANN has the unique superiority of gaining mathematical model aspect from the empirical datum through automatic studying, it does not need the people to assign the form of formula in advance, but it takes the empirical datum as the foundation, and obtain the mathematical model which reflects empirical datum inherent laws after finite iterative computing. The ANN is especially suitable to process the questions with ambiguity rule , many composition and parameters . Therefore, we can use the ANN actualizing direct mapping between the chemical composition and rolling parameters of hot-tolled strip with the mechanical property, then achieve a better prediction of mechanical property of hot–rolled strip.The theory basis and modeling method of the ANN is investigated in this paper. On the base of chemical composition and rolling parameters of hot-rolled Q345B produced by CSP technology in LY Steel, the paper has established the mechanical property prediction model through BP neural network. This model is typical three layers network , namely the input layer, the hidden layer and the output layer. The input layer has 11 neurons, including the content of C, Si, Mn, P, S, Al, Ca and the F1 mill input velocity, the finishing end temperature, the coiling temperature and fished strip thickness. The output layer has 3 neurons, including yield strength, tensile strength and elongation . The prediction result after training of the ANN which compares the data through field surveying indicates that the BP neural network has good precision in mechanical property prediction and is suitable to extend. At the same time, we can investigate between the chemical composition and rolling parameters with the mechanical property and discover laws easily, which provides the important instruction significance and the reference value for the on-line prediction of mechanical property of finished strip and the optimization and the adjustment of product rolling.更多还原