基于神经网络集成的挥发分近红外回归模型

Volatile-NIRS regression model on the basis of neural network ensemble

针对煤炭光谱特征信息分散的现象,提出了基于神经网络集成的挥发分近红外回归模型.该模型引入集成学习的思想,综合SOM,RBF,BP和Elman神经网络学习算法的优势,通过求各子模型的输出均值获得最终的预测结果.为了减小因算法参数设置不当而引起的学习误差,根据各网络算法的特点,利用经验知识、交叉验证和遗传算法优化模型参数.研究结果表明:经相同算法优化后,集成学习模型的性能明显优于单一神经网络,其最大误差小于3%,比单一神经网络小1～2倍.该方法有效地提高了模型的学习精确度,且具有较好的泛化性,适用于复杂多变的非线性煤质近红外回归问题.

Aiming at the information dispersal phenomenon of spectral characteristics, a volatile-NIRS （near-infrared spectroscopy） regression model was presented on the basis of neural network ensemble. The model introduced the ensemble learning method, integrated with the advantage SOM, RBF, BP and Elman neural network, and obtained the final prediction results in the use of calculating the average value of every sub-model＇s output. In order to reduce the errors made by parameters setting in the model, the model parameters was optimized by means of experience knowledge, cross validation and genetic algorithm of each neural network. The results show that： after optimizing the same algorithm, the performance of ensemble learning model is significantly better than that of single neural network sub-models, and the maximum error of ensemble learning model is less than 3 %, and less than 1--2 times in comparison with other single network model. The proposed model effectively improves learning accuracy and has good ability of generalization. Thus, is suitable for complicated, various and nonlinear NIRS regression model of coal quality analysis.