Small-sample machining quality prediction via a fuzzy broad learning system enhanced by prior knowledge

Abstract

Surface roughness is a critical indicator of machined workpiece quality, and accurately modeling its relationship with process parameters is essential for process optimization and intelligent decision-making. Fuzzy broad learning system (FBLS) has demonstrated considerable advantages in nonlinear predictive modeling; however, its performance under small-sample conditions may be limited due to an incomplete rule base and the lack of explicit physical mechanisms. To address this challenge, this article proposes a knowledge-enhanced fuzzy broad learning system (KEFBLS) that integrates dual sources of prior knowledge — expert-knowledge-guided fuzzy partition and physics-based fuzzy rule consequents — to improve predictive accuracy and generalization ability. The effectiveness of KEFBLS is validated on both real-world robotic grinding experiments and a publicly available machining dataset, achieving average prediction errors of only 10.3% and 4.7%, respectively, representing over 20% accuracy improvement over the FBLS baseline. These results highlight the significance of combining domain-specific prior knowledge with data-driven learning, enabling robust performance under limited-data conditions. Overall, KEFBLS provides a unified knowledge- and data-driven framework for surface roughness prediction, with potential applicability to other manufacturing processes where labeled data are scarce.