基于优化Kohonen网络的合金钢Ac1／Ac3单点临界温度预测

Study on Single Point Critical Temperature Prediction of Ac1/Ac3 of Alloy Steel Based on Optimized Kohonen Network

本文利用逐步回归确定Ac1、Ac3转变温度的关键化学元素，进而确定Kohonen神经网络输入层神经元数，优化网络结构；结合无导师学习与监督学习，改进学习算法。基于优化的Kohonen神经网络，建立Ac1／Ac3转变温度预测模型。实验表明，基于Kohonen网络的Ac1／Ac3预测模型的预测精度较高，Ac1相对误差小于3．01％，Ac3相对误差小于3．02％，明显优于逐步回归的预测精度。根据Kohonen网络预测Ac1／Ac3临界温度，进而确定实际加热温度，对钢材热处理保证质量、缩短物理实验周期，具有重要的应用价值。

In this paper, stepwise regression is used to determine the critical chemical elements of Ac1 and Ac3 transition temperature, and then the number of neurons in the input layer of Kohonen neural network is determined, and the network structure is optimized; Combining unsupervised learning with supervised learning, an improved learning algorithm is proposed. Based on the optimized Kohonen neural network, a prediction model of Ac1/Ac3 transition temperature is established. The experimental results show that the prediction accuracy of Ac1/Ac3 model based on Kohonen network is higher, the relative error of Ac1 is less than 3.01%, and the relative error of Ac3 is less than 3.02%, which is obviously better than that of stepwise regression. The prediction of Ac1/Ac3 critical temperature based on Kohonen network and the determination of actual heating temperature has important practical value for the quality assurance of steel heat treatment and shortening the cycle of physical experiment.