Model-Based Prediction of Grinding Surface Roughness With Error Correction via a Knowledge-Based Fuzzy Broad Learning System

Abstract

Surface roughness is a key indicator of product quality. Traditional measurement methods generally rely on manual trials and inspections, which are time-consuming and costly. Therefore, developing an accurate and reliable prediction model is vital for automatic grinding, as it enables more informed decision-making, reduces waste, and improves efficiency. However, current model-driven and data-driven methods suffer from either low accuracy or poor interpretability. To address these challenges, this paper proposes an enhanced hybrid framework that integrates a mechanistic model-based predictor with a knowledge-based Fuzzy Broad Learning System (KBFBLS) for error correction. The mechanistic model offers a physically interpretable baseline estimate of surface roughness, while KBFBLS enhances prediction accuracy by learning the mapping from process parameters to the residual errors of the mechanistic model. Built upon the Fuzzy Broad Learning System (FBLS)—an emerging neuro-fuzzy model for efficient nonlinear modeling—KBFBLS integrates expert knowledge-guided fuzzy partition and variance-determined Gaussian membership function widths, two novel strategies that further improve the system’s expressiveness and adaptability, making it a powerful error corrector. Experimental results on real-world robotic disc grinding tasks show that the proposed framework outperforms representative model-driven, data-driven, and hybrid methods. Furthermore, its adaptability to other machining processes is validated using the wheel grinding and milling datasets.