Gear tooth surface roughness classification under data-scarce conditions using cluster optimal transport-based domain adaptation

Abstract Gear tooth surface precision is critical in high-end manufacturing, yet its inspection remains challenging due to the complex geometry of gears and the scarcity of labeled datasets. Traditional non-contact methods struggle with occlusion and data limitations, while deep learning approaches require large gear-specific datasets that are costly to acquire. To address these issues, this paper proposes a novel domain adaptation framework, the Cluster Optimal Transport-based Dual-branch Multilevel Deep Feature Fusion (COT-DMDFF) model, which enables accurate gear inspection with minimal gear data. The framework first pre-trains a dual-branch network (combining ResNet18 and DeiT-S) on low-cost planar workpiece datasets to capture both local and global surface features. It then employs a cluster-based optimal transport mechanism to align feature distributions between planar and gear domains, minimizing distribution discrepancy without requiring extensive gear samples. Furthermore, a physics-based simulation model is developed to generate realistic gear tooth morphology, expanding the dataset and enhancing model robustness.Experimental results demonstrate that COT-DMDFF achieves an average accuracy of 91.48% in cross-domain tasks, outperforming state-of-the-art methods such as DDAN (81.77%), DSAN (76.09%), and DAN (65.06%). This work offers a cost-effective, reliable, and scalable solution for automated gear precision inspection, with significant potential for applications in aerospace, automotive, and precision manufacturing industries