A Novel Prediction Algorithm for Cigarette Optimizition Parameters with Controllable Tar Amount Based on Invertible Neural Networks

• Published in: IEEE access p. 1
• Main Authors: Jian, Zhou, Sikai, Wang, Fengcheng, Shi, Xiaolin, Zhang, Xiaopeng, Li, Pengcheng, Zhu, Yuanwen, Zou, Xuehui, Sun, Cong, Nie, Yanju, Liu, Xin, Huang
• Format: Journal Article
• Language: English
• Published: IEEE 2024
• Subjects: Agricultural products; Atmospheric modeling; Bidirectional Training; Computational modeling; Couplings; Data models; Invertible Neural Networks (INNs); Jacobian matrices; Multimodal Mapping; Neural networks; Permeability; Predictive models; Production; Tar Content Prediction; Tobacco Production; Training
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3493424

This study proposes a novel approach utilizing Invertible Neural Networks (INNs) to address the complexity of predicting tobacco production parameters from specified tar content in a multimodal mapping task. The INN model takes advantage of bidirectional training and latent variables to accurately capture nonlinear relationships between inputs (Cigarette Paper Air Permeability (CPAP), Tipping Paper Air Permeability (TPAP), and Filter Rod Pressure Drop (FRPD)) and tar content. Experimental results show that the INN model achieves a Mean Normalized Percentage Error (MNPE) of 1.46%, outperforming traditional models like decision trees and linear regression in terms of prediction accuracy. These findings highlight the INN model's potential for precise tar control in tobacco production.