Research on cigarette blend similarity evaluation method based on non-parametric kernel density estimation

ZHAO Zhenjie; LI Shitou; TIAN Yunong; HE Wenmiao; ZHANG Hongfei; PANG Yongqiang; LIAO Fu

doi:10.12187/2025.06.011

Abstract：[Objective]To enhance the efficiency of cigarette blend formulation maintenance,this study proposes a method based on non-parametric kernel density estimation(KDE)for evaluating the similarity of cigarette blend formulations.[Methods]A KDE distribution model for individual chemical components was constructed using historical formulation data to calculate the similarity of individual chemical components.Different weight allocation strategies were employed to compare the similarity between the experimental formulation and historical formulations.The model's performance was evaluated using correlation analysis,and the relationship between the overall similarity of leaf group formulations and sensory quality was analyzed.[Results]The probability density estimation function fitted using KDE effectively reflects the distribution of chemical component data from historical formulations.Compared with the experimental leaf groups,the average overall similarity of the leaf group formulations that passed evaluation to historical formulations in three indicators(sugar-alkaloid ratio,total nitrogen content,and reducing sugar content)increased by 30.57%,18.97%,and 13.06%,respectively.Although the four weighting methods exhibited similar correlation levels,they differed in their tendency regarding the correlation degree of individual chemical component similarity.Specifically,the game theory-based combined weighting method showed correlations with most sensory indicators that were intermediate between those of subjective and objective weighting methods,thereby reducing the overall weight deviation between these two approaches.[Conclusion]Constructed a KDE distribution model using historical formulation data to evaluate the similarity between experimental formulations and historical formulations.The results provide a reference for assessing and controlling the similarity and stability of leaf groups during the formulation maintenance process.