Physics-informed deep learning method for surface roughness prediction in milling process

Abstract

Surface roughness is critical to the functionality and aesthetic performance of mechanical components, necessitating precise prediction and control during the milling process. However, physics-based methods and data-driven methods either exhibit poor performance or lack interpretability, limiting their practical application. To address the issues, this article proposes a novel physics-informed deep learning (PIDL) for milling surface roughness prediction. The core idea is to leverage the principles of cutting mechanics and surface roughness to guide the model construction and regulate the network learning process. Firstly, a high-fidelity dynamic milling force model is established to generate simulated force signals for multi-sensor fusion with other measured signals. Then, the output of the surface roughness physical model is used as physics-guided knowledge to construct an attention-enhanced BiLSTM-BiGRU network based on cross physics-data fusion. In addition, a physics-informed loss function is designed to guide model training, thereby enhancing prediction performance and interpretability. The feasibility and superiority of the proposed method are validated through a series of milling tests under varying conditions. The results indicate that the proposed PIDL can achieve accurate surface roughness prediction in complex milling scenarios, with a coefficient of determination of 0.9845, a root mean square error of 0.0350, a mean absolute percentage error of 1.2895%, and a mean absolute error of 0.0284, outperforming both physics-based methods and data-driven methods.