Search papers, labs, and topics across Lattice.
This paper introduces a hybrid deep learning model, TCN+SKNet+Transformer, to predict tobacco export moisture content by integrating local features, multi-scale information, and global temporal dependencies. The model leverages TCN for capturing temporal dependencies, SK-Net for fusing multi-scale features, and Transformer for modeling global dependencies. Experiments on industrial data demonstrate that the proposed model outperforms traditional methods in terms of mean square error, root mean square error, average absolute error, and R² coefficient, showcasing its effectiveness in multi-dimensional dynamic time series prediction.
A novel TCN+SK-Net+Transformer architecture slashes moisture content prediction errors in tobacco production, paving the way for smarter manufacturing.
Accurate prediction and control of tobacco export moisture content is crucial for improving production efficiency and ensuring the quality and stability of tobacco products, as moisture content is significantly influenced by complex dynamic interactions among multi-dimensional variables such as temperature, humidity, and flow rate. Traditional models fail to meet the demand for high-precision prediction due to their limitations in capturing multi-scale features and long-time dependence. To address these challenges, this study proposes a hybrid model combining a temporal convolutional network (TCN), selective kernel network (SK-Net), and Transformer, which effectively fuses local features, multi-scale information, and global temporal dependencies. The TCN captures short-term and long-term dependencies through inflated convolution and optimizes information transfer via a residual structure; the SK-Net flexibly fuses multi-scale features to enhance key information extraction; and the Transformer models global dependencies using the self-attention mechanism. Data processing involves principal component analysis for feature extraction and normalization to optimize input data quality. Experiments based on industrial refill and recycle equipment data verify the model's superiority. Compared with traditional models, the proposed TCN-Transformer hybrid model significantly reduces mean square error, root mean square error, average absolute error, and increases the R² coefficient, demonstrating strong generalization ability in multi-dimensional dynamic time series prediction tasks. These findings provide important reference for intelligent tobacco production and have significant practical application potential.
