Prediction of Geometric Characteristics of Melt Track Based on Direct Laser Deposition Using M-SVR Algorithm

Abstract

Geometric characteristics provide an important means for characterization of the quality of direct laser deposition. Therefore, improving the accuracy of a prediction model is helpful for improving deposition efficiency and quality. The three main input variables are laser power, scanning speed, and powder-feeding rate, while the width and height of the melt track are used as outputs. By applying a multi-output support vector regression (M-SVR) model based on a radial basis function (RBF), a non-linear model for predicting the geometric features of the melt track is developed. An orthogonal experimental design is used to conduct the experiments, the results of which are chosen randomly as training and testing data sets. On the one hand, compared with single-output support vector regression (S-SVR) modeling, this method reduces the root mean square error of height prediction by 22%, with faster training speed and higher prediction accuracy. On the other hand, compared with a backpropagation (BP) neural network, the average absolute error in width is reduced by 5.5%, with smaller average absolute error and better generalization performance. Therefore, the established model can provide a reference to select direct laser deposition parameters precisely and can improve the deposition efficiency and quality.