Physics-Guided Meta-Learning for Surface Roughness Prediction Under Various Working Conditions With Limited Data

Surface roughness is a key quality metric in manufacturing, making the accurate modeling of its relationship with process parameters crucial for optimizing production decisions. However, under multi-condition settings with limited sample availability, conventional modeling methods often exhibit inadequate accuracy and poor generalization. To address these limitations and accurately predict surface roughness under various working conditions with limited data, this article proposes a physics-guided meta-learning (PGML) method. Surface roughness prediction under different conditions is framed as a few-shot regression problem, with a three-layer fully connected neural network as the regressor. The model-agnostic meta-learning (MAML) algorithm optimizes model initialization, enabling rapid adaptation to new conditions with minimal samples. In addition, physical prior knowledge is integrated within the MAML framework to enhance model training. The effectiveness of PGML is validated through real-world robotic disc grinding experiments, demonstrating superior accuracy and generalization compared to existing methods. Its cross-process applicability is further demonstrated through tests on datasets from wheel grinding and turning. Overall, PGML establishes a novel paradigm that seamlessly combines physical knowledge with meta-learning, offering a possible solution for few-shot surface roughness prediction in manufacturing systems.