基于置信学习机与近红外光谱的煤种快速分类方法  

Rapid Coal Classification Based on Confidence Machine and Near Infrared Spectroscopy

作　　者：王雅圣[1] 杨梦 骆志远 王酉[1] 李光[1] 胡瑞芬[1] 

机构地区：[1]浙江大学智能系统与控制研究所,工业控制技术国家重点实验室,浙江杭州310027 [2]Computer Learning Research Centre,Royal Holloway,University of London,Egham Hill,Egham,Surrey TW20 0EX,UK

出　　处：《光谱学与光谱分析》2016年第6期1685-1689,共5页Spectroscopy and Spectral Analysis

基　　金：国家高技术研究发展技划项目(2013AA041201);浙江省科技计划项目(2015C37062)资助

摘　　要：基于近红外光漫反射谱技术的检测分析具有简单,快捷,安全等优势而被广泛应用于各行各业。应用近红外光谱分析技术实现不同煤种的快速分类,该方法可以替代费时费力费财的传统化学分析方法。同时首次将置信学习机(confidence machine)引入近红外分析中,实现了对分析结果的风险评估。采集了来自不同矿区共四种不同煤种(肥煤,焦煤,瘦煤和贫瘦煤)的199个煤样本的近红外光谱,通过机器学习的方法针对煤的近红外光谱构建了煤种分类器来实现煤种的快速分类。在近红外分析中引入了置信学习机的分析方式,结合支持向量机(SVM),构建了离线和在线的CM-SVM分类器。置信学习机是一种概率方法,使用概率(CM-SVM)来取代分类超平面(SVM)进行分类,不仅分类效果好于传统的SVM,达到了95.48%的分类率,还能同时给出每个样本分类结果的置信度,可靠度等风险信息。另外,CM-SVM通过对置信水平的设定,得到不同置信度下预测区间,该区间的预测正确率是与置信水平严格对应的,对于产品质量控制有非常重要的意义。置信学习机同时是一种在线的学习模型,新样本的不断加入会提高模型的性能,非常适合于工业现场的在线分析。在线的CM-SVM模型随着样本数的增加,预测结果的置信度有所提高,对工业现场近红外分析有重要意义。Near-infrared reflectance spectroscopy （NIRS） is a simple, convenient and safe technology which is widely used in many industries. NIRS was employed to the rapid classification of coal in this study. The new method can be a replacement of the chemical analysis which is laborious and time consuming. Confidence machine was firstly applied to NIRS in this study which was used to evaluate the risk of the analysis. The near infrared reflectance spectrum of 199 coal samples including four types of coal （50 fat coal samples, 50 coking coal samples, 49 lean coal samples and 50 meager lean coal samples） from different mines in China were collected and classifiers based on the near infrared spectra of coal samples which were established by using machine learning methods to realize the rapid classification of coal samples. Confidence machine was introduced into the analysis technology based on NIRS in this paper. Confidence machine based on support vector machine （CM-SVM） was built and applied to the classification of coal samples via NIRS. Confidence machine is a probabilistic algorithm and instead of using hyper plane （SVM） to carry out the classification, using probability （CM-SVM） turned to be more effective which had 95.45 % of the samples correctly grouped. Besides that, CM-SVM also estimated the confidence and credibility for each predicted sample. By setting different confidence levels, CM-SVM can perform region prediction whose error rate was predefined by the different confidence levels, which was very important for the control of product quality when NIRS was applied to the analysis of productions. Confidence machine is designed as an on-line learriing methodl new samples can be added to the training set one by one to improve the efficiency of the model and is very appropriate for industry on-line analysis. On-line CM-SVM models showed that the confidence of prediction would be raised as the samples increased, which was valuable for industry on-line analysis.

关 键 词：近红外光谱 煤种分类 置信学习机 支持向量机 

分 类 号：O657.3[理学—分析化学]