Abstract

The moisture content of tobacco, as an important characteristic which should be kept at a desired level to maintain consistent product quality in drying process, is difficult to perform the direct measurement and anomaly detection due to its large delay in actual process. Therefore, an intelligent real-time detection method is an urgent and challenging task in ensuring the product quality. This paper proposes a time-domain raw data conversion method along with a novel deep learning architecture called multi-hierarchical convolutional neural network (MHCNN) for moisture prediction, in which the proposed architecture automatically learns multi-hierarchical features from transformed image-like data and simultaneously performs online prediction. Experiments are conducted on the real production data from the cigarette factory and the presented model performs well on overall testing dataset. Specifically, the MAE, RMSE and R² of normal production batch can reach to 0.0131, 0.0244, and 0.9721 respectively, which are far superior to the estimation of experience and other alternatives. It demonstrates that the proposed online prediction strategy can simultaneously perform multi-hierarchical feature extraction and moisture online prediction with high precise to eliminate the detection delay for process optimization and control.

摘要

烟草的水分含量作为在干燥过程中保持产品质量一致的重要特性，由于其在实际过程中的较大延迟，难以进行直接测量和异常检测。因此，智能实时检测方法是保证产品质量的一项紧迫而具有挑战性的任务。本文提出了一种时域原始数据转换方法以及一种称为多层次卷积神经网络 (MHCNN) 的新型深度学习架构，用于水分预测，其中所提出的架构自动从转换后的类图像数据中学习多层次特征，并同时进行在线预测。对卷烟厂的真实生产数据进行了实验，所提出的模型在整体测试数据集上表现良好。具体来说，MAE、RMSE 和正常生产批次的R ²可分别达到0.0131、0.0244和0.9721，远远优于经验估计和其他替代方案。表明所提出的在线预测策略可以同时进行多层次特征提取和水分在线预测，精度高，消除了工艺优化和控制的检测延迟。