Machine Learning-Based Surface Roughness Prediction of Drilled Hole Surfaces Using Texture and Chip Morphology Features

Abstract

Surface roughness ( ) is a significant quality indicator in machined components, influencing performance, fatigue life, and tribological behaviour. Conventional contact-based stylus profilometers are slow, risk micro-scratches, and are impractical for real-time measurement of the surfaces of internally drilled holes. This study proposes a Machine Learning (ML)-based framework that estimates  from image-derived texture features, machining parameters, and chip morphology descriptors—specifically addressing drilled hole interiors where curvature and non-uniform illumination render standard planar analysis ineffective. Three models were evaluated: (i) Linear Regression (LR) with GLCM texture features (  = 0.7621); (ii) Extreme Gradient Boosting (XGBoost) combining image texture features with spindle speed and feed rate (  = 0.8201, MAE=0.1425 µm); and (iii) XGBoost using chip morphology features with chip area and curl index extracted from chip images (  = 0.8092). A dataset of 370 drilled-hole images of Al-7075 alloy was augmented to 1800 samples. One-way ANOVA confirmed statistically significant performance differences between models (F (2,537) = 47.3, p<0.001). A Streamlit-based interface enables real-time, non-contact  prediction at ~45 ms per image, reducing inspection time, cost, and reliance on contact methods.