Z比分数和聚类分析在烟草化学成分检测能力验证中的应用

Z-score and Cluster Analysis in Assessment of Comprehensive Detection Ability of Tobacco Chemical Composition

为探究不同Z比分数法对单个指标评价分析的异同,解决Z比分数法无法满足多指标综合评价的问题。本文以烟叶常规化学成分检测的能力验证为场景,一方面比较常规Z比分数法和稳健Z比分数法在单个指标检测能力评定的异同性,优选适宜单指标评价方法;另一方面将Z比分数、主成分分析、聚类分析相结合,建立多指标综合评价方法。评价结果表明:(1)在单指标能力评价中,常规Z比分数和稳健Z比分数求得的实验室满意率均达85%以上,符合经验预期;(2)两种Z比分数的数值相关系数达99.90%以上,且跟随性较好;(3)稳健Z比分数计算更方便,且在不同数据分布形态下具有更好适用性,而常规Z比分数在数据服从或近似正态时有较高适用性;(4)常规Z比分数与聚类分析相结合能有效将40个实验室分为6类,但第一类实验室占比达80%,类与类之间分离度不理想;(5)常规Z比分数先主成分提取再聚类能提高类与类间分离程度和类特征鲜明性,其中第一,二类实验室代表能力验证满意,且检测结果系统性略偏低或略偏高的实验室,第三,四、五、六类则分别代表某一指标不满意、水溶性总糖明显偏低、总植物碱明显偏高、总氮明显偏高的实验室。

In order to explore the similarities and differences in evaluation of Single Index by different Z-Score methods,and to solve the problem that the Z-Score cannot satisfy the comprehensive evaluation of multiple indicators.This article takes the verification of the ability to detect the conventional chemical composition of tobacco leaves as a scenario.On one hand,it compares the similarities and differences between the conventional Z-Score and the robust Z-Score in the single index detection ability evaluation,and to evaluate which one is preferred;on the other hand,the combination of Z-Score and principal component analysis and cluster analysis establishes a multi-index comprehensive evaluation method.The evaluation results show that:(1)In the single-index ability evaluation,the laboratory satisfaction rate obtained by the conventional Z-Score and the robust Z-Score is over 85%;(2)The correlation coefficient of two Z-Score is more than 99.90%,and the followability is better;(3)The robust Z-Score is more convenient to calculate and has better applicability under different data distribution forms,while the conventional Z-Score has high applicability when the data comply with abnormal distribution;(4)The combination of Z-Score and cluster analysis can effectively divide 40 laboratories into 6 classes,but more than 80%of laboratories are included in first class;(5)Conventional Z ratio scores with principal component extraction followed by clustering can improve the degree of class-to-class separation and the distinctiveness of class characteristics.Such as,the first and second class are satisfied with the ability verification,but slightly low or high,the third,fourth,fifth,and sixth class represent an unsatisfactory index respectively.