Physics-Guided Data Augmentation with Nonlinear Supervised Representation Learning for Surface Roughness Prediction in Machining

Abstract Accurate surface roughness prediction in machining is often hampered by data scarcity. While Generative Adversarial Networks (GANs) offer a promising avenue for data augmentation, the synthesized samples often suffer from a lack of authenticity due to the absence of governing physical constraints. To address this, a novel framework, integrating physics-guided domain knowledge with nonlinear representation learning, is proposed. The framework is composed of two core modules: a Physically Constrained GAN (PCGAN) and a Nonlinear Supervised Representation Learning (NSRL) unit. Specifically, PCGAN encodes domain knowledge, derived from process design rules as prior guidance and mechanism-driven statistical priors, into the loss function. This design regularizes the generative process within a physically feasible domain, ensuring synthesized data is both statistically realistic and physically plausible. Concurrently, NSRL learns a task-optimized latent space for effective feature fusion and labeling. Validated on a proprietary industrial grinding dataset HTSPG and the public GAMHE 5.0 macro-milling dataset, the proposed framework significantly outperforms state-of-the-art methods. Experimental results demonstrate that NSRL-PCGAN reduces the Root Mean Square Error (RMSE) by 6.8% and 31.5% respectively, proving its superior generalization capability in complex manufacturing scenarios.