Evolutionary algorithm-optimized feature fusion for accurate classification of shredded tobacco using multi-sensor data

Long Chen^1,2Ni Tang¹Xiao Wu¹Yang Wang¹Chuan He¹Zongwei He¹Lihua Xie¹Xixiang Zhang¹Xing Chen²Tao Zhou^1*

• ¹China Tobacco Sichuan Industrial Co Ltd, Chengdu, China
• ²Biosensor National Special Laboratory Department of Biomedical Engineering, Zhejiang University, Hangzhou 310027, China., Zhejiang University, Hangzhou, China

This study presents a novel evolutionary algorithm-based feature fusion framework for enhancing sensing accuracy in shredded tobacco classification. To overcome the limitations of individual-sensor systems, data from GC-SAW, E-nose, and FTIR were synergistically fused. A systematic comparison confirmed feature-level fusion as the most effective strategy, yet its performance depended on overcoming the high dimensionality of the fused data. We rigorously evaluated seven dimensionality reduction methods, and the genetic algorithm (GA) was identified as the cornerstone of our framework. The GA-based feature selection demonstrated exceptional performance, achieving a mean classification accuracy of 99.89% ± 0.79% across 50 independent test runs by intelligently distilling the fused feature set into a compact, highly discriminative subset that balanced information from three sensing modalities to maximize their complementary strengths. Our work confirms that evolutionary algorithm-based feature fusion is a powerful and robust method for unlocking the full potential of multi-sensor data, thereby significantly advancing the accuracy of complex plant material classification.