密集烤房烟叶含水率空间分布在线检测系统设计与试验

Design and Experiment of Online Detection System for Spatial Distribution of Tobacco Leaf Moisture Content in Bulk Curing Barns

摘要:

密集烤房烘烤过程缺乏含水率空间分布在线检测手段,导致工艺调控响应滞后,烟叶烤青、挂灰等问题频发。 提出了一种基于稀疏称量网络与机器学习的烟叶含水率空间分布在线检测新方法。该方法以烤房前端截面的烟叶质量、瞬时脱水速率及环境温湿度作为动态特征参量,结合目标点的空间坐标,建立了从含水率局部测量到空间分布的非线性映射关系。在此基础上,采用贝叶斯优化(BO)对构建的XGBoost 模型超参数进行全局优化,获得适应阶段性变温调湿特性的含水率空间分布预测模型,并集成称量传感网络开发在线检测系统,进行了验证试验。 结果表明,测试集中预测模型的R2、MAE和RMSE分别为0. 996、0. 76% 和1. 14% ,预测性能优于SVR、MLP对比模型。在线验证试验发现,烟叶含水率预测值与实测值的平均R2、MAE 和RMSE分别为0. 978、2. 77% 和3. 66% ,表明检测系统的可靠性。变黄期(38℃)、定色期(48℃)和干筋期(65℃)含水率空间分布的变异系数分别为1. 73% 、 8. 27% 和20. 02% ,明确烟叶含水率均匀度随烘烤进程逐步下降,证实检测系统可对含水率空间分布均匀性进行有效量化。研究结果为实现烟叶烘烤工艺参数的精准调控提供了技术支撑。

Abstract:

The lack of online detection methods for spatial moisture content distribution during the curing process in bulk curing barns leads to delayed process control responses and frequent issues such as greenish and spotted tobacco leaves. A novel online detection method was proposed based on a sparse weighing network and machine learning. This method established a nonlinear mapping from local moisture content measurements to spatial distribution, using dynamic features including tobacco weight, instantaneous dehydration rate, and ambient temperature and humidity from the front section of the barn, combined with the spatial coordinates of target points. On this basis, Bayesian optimization (BO) was employed to globally optimize the hyperparameters of the constructed XGBoost model, resulting in a predictive model adapted to the stage-varying temperature and humidity characteristics. An online detection system integrated with the weighing sensor network was developed and validated. Results demonstrated that the proposed prediction model delivered excellent prediction performance on the independent test set, with a coefficient of determination (R2) of 0. 996, a mean absolute error (MAE) of 0. 76% and a root mean square error (RMSE) of 1. 14% for the original wet basis moisture content of tobacco leaves, significantly outperforming the conventional SVR and MLP models. Online validation tests showed that the average R2 between the predicted and measured original moisture content was 0. 978, with mean absolute error of 2. 77% and average RMSE of 3. 66% , demonstrating the system??s high reliability. The coefficients of variation for spatial moisture content distribution during the yellowing (38℃), color-fixing (48℃), and stem-drying (65℃) stages were 1. 73% , 8. 27% , and 20. 02% , respectively, clearly indicating a gradual decline in moisture uniformity as curing progressed and confirming the system??s capability to effectively quantify the spatial distribution uniformity. The findings can provide technical support for the precise control of tobacco curing process parameters.