选择性神经网络二次集成在火药近红外分析中的应用研究

The Two-Level Selective Neural Network Ensembles Applied to Quantitative Analysis of Propellants

针对建立近红外光谱定量分析的神经网络校正模型时,存在变量数过多以及容易出现过拟合等问题,采用小波变换对近红外光谱进行预处理,用以消除噪声,减少变量个数;并在此基础上,提出一种新的神经网络校正模型—基于改进贪心法的选择性神经网络二次集成,来提高神经网络的泛化能力。实验结果表明:在建立火药近红外分析的校正模型中,该模型不仅建立过程简单而且具有较好的泛化能力。

During quantitative analysis of propellant based on near infrared spectroscopy （NIR）, neural networks are good tools for establishing calibration model, but there are still some problems to be considered such as too much variables and ＂over-fitting＂. To solve these problems, wavelet transform （WT） was used to preprocess the NIR, and the two-level selective neural network ensembles based on ＂greedy algorithm with first improvement strategy＂, were proposed. Based on the ensembles, a calibration model for the quantitative analysis of propellant was established. Experiment results show that the model is easy to be built and promotes the generalization ability of neural network system.