木材导热系数非线性拟合的神经网络模型

Nonlinear fitting calculation of wood thermal conductivity using neural networks

为了得到木材导热系数随各种物性参数变化的非线性规律,提出了一种基于人工神经网络的预测计算模型.该模型以木材的温度和孔隙率为输入量,以其导热系数为输出量.设隐层的神经元数分别为3～8,构建了6个不同隐层结构的神经网络,并利用桦木的导热系数分别对这6个神经网络进行了训练.通过对误差的比较分析,得到了具有最优性能的网络型式,其隐层具有6个神经元,训练结果的平均相对误差为0.21%,平均绝对误差为4.33×10^-4 W/（m·K）.利用该最优网络对不同温度和孔隙率下桦木的导热系数进行了预测,其结果与已知的实验测量数据吻合得很好,表明神经网络理论可以有效地用于木材导热系数的非线性拟合计算而且具有较高的精度.

A computational model based on artificial neural networks （ANN） was proposed to find the nonlinear variance regulation of the thermal conductivity versus physical properties of wood. The temperature and the porosity of wood were set as the two inputs and the thermal conductivity of wood was set as the output. The number of neurons within the hidden layer varied from three to eight with increment of one, resulting in a total of six networks. The thermal conductivity of birch was predicted using these six networks respectively. The optimal network was recognized as the one which had six neurons within the hidden layer by comparison and analysis of the errors. The mean relative error was 0.21%, and the mean absolute error was 4.33×10^-4 W/（m·K）. This network was used to predict thermal conductivities of birch at different temperature and porosity. Results demonstrated good agreement between the predicted and the available experimental data, which showed that the ANN model can be used to effectively predict the thermal conductivity of wood and that it has ideal accuracy.