PhyDT: Physics-Enhanced Data-Driven Transformer for Cutting Force Prediction of Boring Process

The prediction of cutting force in boring machining is crucial for understanding cutting deformation, tool wear, and work-piece quality. Physics-based models offer adaptability across various working conditions but struggle with complex dynamic scenarios and real-time effectiveness. Conversely, deep learning models, while effective in capturing nonlinear relationships, are highly data-dependent and may lose accuracy in unfamiliar conditions. This study proposes a physics-enhanced data-driven transformer (PhyDT) model for real-time cutting force prediction, integrating physics knowledge with deep learning. Specifically, the PhyDT leverage the information of both machining parameters and theoretical cutting force (calculated using a modified model based on Merchant theory and Johnson–Cook constitutive model) into a transformer network with time-series sensory data as input. By incorporating these physics-based insights, the model aims to balance adaptability and accuracy across different machining conditions, offering improved performance in complex scenarios. We designed and built a heterogeneous data acquisition platform for boring processes to validate the PhyDT model under various experimental parameters. Results show improved prediction accuracy and generalization on an untrained workpiece compared to traditional data-driven methods. An ablation study shows that incorporating machining physics information and multiscale attention cascaded with convolutional neural networks improves performance, reducing mean absolute percentage error to 0.623% .