An explainable AI-guided feature refinement framework for surface roughness prediction in robotic drilling

Abstract

In aircraft assembly, the low structural stiffness of industrial robots complicates surface quality control, and the opacity of conventional machine learning models hinders their adoption for process optimization. To address this, this paper presents a systematic Explainable AI-Guided Feature Refinement Framework to develop a minimal, yet robust, and physically interpretable model for surface roughness prediction in robotic drilling. The framework utilizes SHapley Additive exPlanations (SHAP) as an active component in an iterative feature selection process to refine a Random Forest model. The experimental validation on an integrated industrial platform demonstrates that this approach successfully identifies a minimal set of critical features from a high-dimensional dataset, including process parameters and specific vibration characteristics. The resulting model achieves superior predictive performance and stability compared to conventional feature selection methods. Furthermore, the analysis uncovers key non-linear relationships and feature interactions, providing interpretable insights into how operational parameters and dynamic responses collectively influence surface quality, which facilitates process optimization in robotic aircraft assembly.