Paper Quality Enhancement and Prediction Using Deep Learning Architectures

Abstract

This paper proposes an enhanced approach to predicting paper quality parameters using advanced deep learning models. In contrast to the state of art studies that employed traditional machine learning (with k-Nearest Neighbors as the best predictor), a Convolutional Neural Network (CNN), Long Short- Term Memory (LSTM) network, a hybrid CNN-LSTM, and a Transformer-based model were used. A realistic paper manufacturing dataset is stimulated in this work with key variables (such as moisture, grammage, caliper, dryer temperature, ambient humidity, pulp composition, and machine speed) and detailed mathematical formulations for each model are provided. Experimental results demonstrate that deep learning significantly outperforms previous methods, the Transformer model achieves a root mean squared error (RMSE) as low as 0.5 (improving upon 2.0 from the best traditional model) and R² above 0.99. Moreover, we introduce interpretability analyses using Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) to explain the model predictions. This interpretable deep learning framework yields highly accurate predictions of paper quality in real time, enabling more efficient control of the drying process and reduction of steam usage while maintaining product quality.