Closed-Loop Parameter Optimization for Robotic Machining Using Physics-Informed Machine Learning and Multiobjective Optimization

Abstract

In practical applications, the simultaneous optimization of numerous design parameters in time-consuming multi-objective optimization experiments is recognized as a significant bottleneck across various scientific and engineering disciplines. A prominent example is the optimization of machining parameters for achieving efficient and precise robotic belt grinding (RBG). This paper presents a closed-loop machining parameter optimization approach, which comprises two key stages: forward multi-task prediction and backward multi-objective parameter optimization. In the first stage, a physics-informed neural network (PINN) method is introduced, which integrates the multi-gate mixture-of-experts multi-task learning method with an RBG mechanism model to simultaneously predict material removal depth and averaged surface roughness. In the second stage, a powered multi-objective particle swarm optimization (MOPSO) method is developed, which combines a standard MOPSO method with a non-linear Powerball technique, to efficiently optimize the RBG machining parameters with a limited number of training iterations based on the learned PINN model. Two optimal machining parameter solutions are generated and recommended for the RBG machining process. The effectiveness and superiority of the proposed closed-loop parameter optimization method are validated through comparative experiments, which demonstrate its advantages in both coprediction accuracy and optimization efficiency. Note to Practitioners—This paper addresses the challenge of identifying robotic machining parameters that effectively balance machining efficiency and surface quality. Traditional trial-and-error methods for adjusting these parameters are both time-intensive and costly, given the vast number of possible combinations. To overcome these limitations, this paper proposes an intelligent optimization approach that leverages historical machining data to automatically determine optimal machining parameters. Our approach integrates artificial intelligence techniques with robotic machining mechanisms, ultimately recommending two sets of parameters for robotic machining. Preliminary experiments demonstrate the feasibility of this approach, but it has not yet been incorporated into a robotic machining system or tested in production. Future research will focus on dynamically optimizing robotic machining parameters by combining dynamic time-series signals with static machining parameters.