Transfer Learning Enhanced Gaussian Process Model for Surface Roughness Prediction with Small Data

Abstract

Surface roughness is critical for the functionality (e.g., wear and fatigue) of a machine component. Nevertheless, the accuracy and efficiency are limited for analytical and simulation methods to predict surface roughness. On the other hand, data-driven methods (e.g., machine learning) provide an alternative for this purpose. However, the power of machine learning models is often limited by the availability of machining data. Therefore, this paper proposed a new model by integrating Transfer Learning (TL) and Gaussian Process (GP) for roughness prediction. The datasets from the literature are leveraged with additional measured data to improve the model accuracy. The model performance is benchmarked with the non-augmented Gaussian Process Model (GPM) and the results show that the TL-enhanced model performs better than the non-enhanced model.