近红外光谱灰分预测模型中煤炭样本的优化方法

Optimization Method of Coal Sample in Ash Prediction Model Based on Near Infrared Spectroscopy

针对近红外光谱灰分预测模型中样本数据特有的问题,首先采用主成分分析方法剔除建模样本集中的异常样本,并提取出煤炭光谱的特征信息;然后提出一种集成自组织映射神经网络和模糊C均值聚类算法的双层聚类方法,将样本集分为5个子集,并滤除其中的争议点;最后搭建基于GA-BP神经网络的煤炭灰分预测子模型,单独分析各子集的测试集样本。实验结果表明,基于主成分分析和双层聚类方法的煤炭样本优化方法不仅能准确排除异常样本和可疑样本,还能有效地压缩样本数据,使得各子模型的学习精度和运算速度得到显著提高。该方法为近红外光谱煤质分析技术的发展应用提供了一种有效可行的新途径。

According to the unique problem of sample data in ash prediction model based on near infrared spectroscopy, an optimization method was proposed. Principal component analysis method is used for eliminating abnormal samples in coal sample set and extracting feature information of coal spectrum. A double-level clustering method is presented which integrates self-organize map neural network and fuzzy C- means clustering algorithm. The method classifies original sample set into 5 subsets and filtered dispute points. At last, prediction sub-models of coal ash are built for each subset based on GA-BP neural network to analyze testing samples of each subsets separately. The experimental results showed that the optimization method based on principal component analysis and the double-level clustering method can check and remove abnormal and suspicious samples exactly, compress sample data effectively, and improve learning precision and calculating speed of sub-models dramatically. The optimization method was a new effective method for development and application of near infrared spectroscopy in coal quality analysis.