Knowledge-wrapping method for prediction and evaluation of material removal behavior in robotic belt grinding

Abstract

Reliable prediction and evaluation of material removal (MR) is a continuing pursuit in the grinding process. Mechanistic and empirical MR models always suffer from inaccuracies and restricted applicability, whereas data-driven approaches remain deficient in sample dependence, generalization, and physical interpretability. This motivates us to develop a knowledge-wrapping method (KWM) to predict and characterize the material removal behavior in robotic belt grinding. A physical material removal profile model (PHY) that suits the robotic belt grinding is first presented by appealing to the Archard law and Hertzian theory. Next, the knowledge-wrapping matrix is designed to wrap the physical constraints into a form of design matrix by transforming the mechanistic and empirical models into a linear system. The knowledge-wrapping matrix and the experimental data are then connected via a likelihood function with the consideration of measuring noise, yielding a hybrid-driven learning process that preserves interpretability from PHY. Comparative experiments and MR mechanism interpretations are finally presented to demonstrate the effectiveness and superiority of the proposed method.