In the era of digital transformation, Remaining Useful Life (RUL) prediction plays a critical role in advancing smart manufacturing and equipment health management. Deep learning, which automatically extracts high-quality features from raw data and reduces reliance on manual feature engineering, has emerged as a primary focus in RUL prediction research. Traditional deep learning models have shown strong predictive performance in manufacturing RUL tasks; however, they often have complex architectures and a large number of parameters. To address this, we propose an adaptive temperature regulation method for soft-target distillation. We introduced a combination of Loss-Based Smoothing and Exponential Decay Based on Training Progress to achieve more flexible and stable temperature control. The goal is to compress the model size while enhancing the student model’s learning efficiency and prediction accuracy. The proposed method is evaluated on the benchmark C-MAPSS dataset using RMSE and Score as the primary evaluation metrics. Experimental results show that our method outperforms both the non-distillation baseline and single-strategy variants across all C-MAPSS sub-datasets. The proposed approach significantly enhances model compactness without compromising prediction accuracy, offering an efficient and practical solution for edge predictive maintenance in real-world industrial scenarios.