Authors

Abstract

The curing of tobacco is a critical process that determines the quality of the final product. Accurate recognition of tobacco curing states is essential for ensuring optimal quality. Existing recognition models mostly focus on the transient states within the curing barn. In contrast, this study incorporates multiple time steps to capture dynamic feature changes in the curing barn over time, providing a more accurate state recognition. A hybrid deep learning model combining Temporal Convolutional Networks (TCN), Long Short-Term Memory (LSTM) networks, and a novel Density-aware Channel Redistribution Unit (DCRU) based on Kernel Density Estimation is proposed. The model integrates the global feature extraction capability of TCN, the long-term dependency modeling strength of LSTM, and the complex channel feature extraction ability of DCRU, thereby enhancing the model's performance in recognizing the stages of tobacco leaf curing. Tests conducted on a real-world tobacco dataset demonstrate that the model achieves a prediction accuracy of 0.989 and outperforms baseline models as well as existing tobacco curing state recognition methods. These results validate the effectiveness of the hybrid TCN-LSTM model in recognizing tobacco leaf curing states, with promising applications in agricultural automation.

Abstract in Chinese

烟叶烘烤是决定最终产品质量的关键工序，准确识别烟叶烘烤状态对于保障其品质至关重要。现有的识别模型多侧重于烘烤过程中烤房内的瞬时状态，而本研究通过引入多个时间步，捕捉烤房内动态特征的变化，从而实现更为准确的状态识别。本文提出了一种融合Temporal Convolutional Network（TCN）、Long Short-Term Memory（LSTM）与基于核密度估计的Density-aware Channel Redistribution Unit（DCRU）的混合深度学习模型。该模型结合了TCN的全局特征提取能力、LSTM对长期依赖的建模能力，以及DCRU在复杂通道特征分布提取方面的优势，从而有效提升了对烟叶烘烤阶段的识别性能。在真实烟叶数据集上的测试结果表明，该模型的预测准确率达到0.989，优于基线模型及现有的烟叶烘烤状态识别方法。研究结果验证了该混合TCN-LSTM模型在烟叶烘烤状态识别中的有效性，为农业自动化应用提供了有前景的解决方案。