Parameter automatic optimization strategy for laser powder bed fusion using neural network infrared radiation intensity prediction model

Abstract

Conventional scanning approaches with fixed laser parameters are prone to causing localized defects and in-homogeneous performance when forming parts with complicated geometries using the laser powder bed fusion (LPBF) process. While online monitoring of the forming process using thermal cameras allows the identification of abnormal infrared radiation intensity values of the scanning line to pinpoint regions of potential defects, it fails to proactively prevent defects from occurring. To solve this problem, this study constructed a back propagation (BP) neural network to predict the infrared radiation intensity of the current layer before scanning. In the meantime, this study designed a parameter optimization algorithm to adjust the local scanning parameters with a view to avoiding defects and thus enhancing the performance consistency of the formed parts. This study drew on the idea of the divide-and-conquer method to divide the parts into multiple units for data acquisition by thermal camera, and has collected 500,000 pieces of experimental data through 15 experiments. By utilizing these data, this study trained a BP neural network to predict the infrared radiation intensity, and its mean absolute error (MAE) of the prediction results was 25, the coefficient of determination (R2) was 0.708, and the F1 score (abnormal radiation intensity identification and classification) was 0.802. Subsequently, the study optimized the scanning parameters of the unit by using particle swarm optimization (PSO), and performed practical application and validation on spherical gear parts of 316 L stainless steel and blade parts of DZ125 nickel-based superalloy. Findings indicated that the automatic optimization strategy can effectively prevent abnormal infrared radiation intensity values during the scanning process and enhance the performance consistency of the LPBF process for complicated parts in both different materials and structures.