基于贝叶斯决策的近红外光谱药片分类方法  

Tablets Classification Method Based on Bayesian Decision and Near Infrared Spectroscopy

作　　者：周扬[1,2] 吕进[1] 戴曙光[2] 刘铁兵[1] 施秧[1] 葛丁飞[1] 李博斌[3] 

机构地区：[1]浙江科技学院近红外应用技术研究室,杭州310023 [2]上海理工大学光电信息与计算机工程学院,上海200093 [3]国家黄酒产品质量监督检验中心,绍兴312071

出　　处：《分析化学》2013年第2期293-296,共4页Chinese Journal of Analytical Chemistry

基　　金：国家自然科学基金项目(No.51075280);浙江省重大科技专项和优先主题计划项目(No.2010C11060);浙江省自然科学基金项目(Nos.Y1100219;Y4110235);浙江省质监系统2011年度科研计划项目(No.20110234)资助

摘　　要：针对药片近红外光谱法分类过程中校正集样本数量过少且各类样本数量不均导致分类误差问题,提出了基于贝叶斯决策的分类方法。本方法对校正集样本在各类中的先验概率密度和各类药片光谱的类条件概率密度进行了估计,利用贝叶斯全概率公式计算了待分类光谱分属于各类的后验概率,根据后验概率大小对药片分类。实验随机选取4类数量不等的西酞普兰药片70片,建立贝叶斯决策分类模型,对20片验证集药片进行分类,各类的分类灵敏度和特异度均达到了100%,对比判别最小二乘法的分类结果,验证了贝叶斯决策分类法能将样本及其近红外光谱的分布信息参与分类决策,提高了分类的准确性和适应性。Lack of samples and uneven distribution in different sample sets is the problem at tablets classification using near infrared spectroscopy that cause the classification error. A classification method based on Bayesian decision making is proposed. The method estimate the priori probability density of calibration set and class conditional probability density of unknown tablets spectral using different kinds of calibration set. The posterior probability is calculated by Bayesian full probability formula. The unknown tablet is classified according to its posterior probability. The experiment selects four categories of citalopram tablets with different amount in each category. The Bayesian decision model based on calibration set of 70 spectra is used to classify 20 tablets in the validation set. Comparing to PLSDA method, the sensitivity and specificity of the Bayesian model is 100%, validating the Bayesian decision taxonomy can improve the classification accuracy and adaptability in the distribution based on NIR spectral.

关 键 词：药片分类 近红外 贝叶斯决策 参数估计 

分 类 号：R927[医药卫生—药学] O657.33[理学—分析化学]