Adaptive online learning for system identification and in-advance optimization under long feedback delay and concept drift in cigarette production

Abstract

The moisture content of tobacco at the end of a drying process is controlled to guarantee the product quality. However, the unforeseen moisture at dryer inlet usually leads to improper dehydration level that degrades the product quality. To overcome the problem, a suitable control reference of moisture should be set much earlier for the conditioning-casing process to reach. This boils down to the challenging task of modeling intermediate moisture dissipation with long feedback delay and streaming data under concept drift. In this paper, a novel method is proposed based on Recursive Least Squares (RLS) for online moisture modeling and in-advance optimization. For modeling, a multi-step representation is employed on processed mechanism-based features to alleviate the feedback delay. Besides, a variable forgetting factor is designed for RLS to maintain the tracking ability toward unpredictable concept drift. Then the optimization can be fast achieved via a feature forecasting strategy and reverse model inference. Extensive experiments are performed on two years’ real production data covering 2280 valid tobacco batches. The algorithmic evaluation shows the proposed method outperforms others on all metrics involved, achieving $R^2$ of 0.815 and 0.875 for year 2021 and 2020 with least single estimation time of 0.213 ms. The on-site evaluation manifests the improvement of production quality represented by the increased stability of drying dehydration level around the standard 5.9%.

Keywords:

• Long delay
• concept drift
• recursive least squares
• multi-step representation
• variable forgetting factor

Additional information

Funding

This work was supported by the Hongta Tobacco (Group) Co., Ltd. for its real production data as well as the Advanced Intelligent Maintenance Systems (AIMS) Co., Ltd. for its technical contribution without specific financial support.