Abstract: In vision based industrial robot repair welding, fast and accurate prediction of material removal depth (MRD) is a key challenge to achieve process intelligence and efficient control. Traditional models often face the dilemma of balancing accuracy and real-time performance. Based on quadratic polynomials and specific grinding energy, this paper establishes a basic prediction model and proposes a lightweight model called Light-RF, a residual optimization approach based on the traditional random forest (RF) machine learning algorithm. This method systematically constrains the model structure parameters, aiming to significantly reduce the computational complexity and volume of the model while maintaining comparable prediction accuracy to traditional RF. The R² is as high as 0.9956, the model size is reduced by 60.17%, and the calculation speed is improved by 55.02% to meet the timely response requirements of industrial sites.