Cross-Line Transfer of Initial Juice Sugar Content Prediction Model for Sugarcane Milling

Some sugar factories lack sufficient data collection in their milling production lines, making it challenging to construct data-driven models for predicting initial juice sugar content. This paper proposes an adversarial semi-supervised pre-training and fine-tuning modeling method. The model is first trained on data-rich source production lines (190,000 samples, 500 labels, 1 min sampling for process variables, 3 times/day for sugar content) and then fine-tuned using limited data from the target production line (10,000 samples, 100 labels), effectively utilizing both labeled and unlabeled data to enhance the model’s generalization ability. Ablation experiments were conducted using data from two sugarcane milling production lines. The experimental results validate the effectiveness of each component of the proposed method and its prediction accuracy, achieving a reduction in MSE by 0.067 (23.0%) and MAE by 0.066 (15.7%) compared to the standard pre-trained model. This strategy not only optimizes the prediction of initial juice sugar content for production lines with insufficient data collection but also has the potential to improve the efficiency and quality of sugarcane milling industrial production.