Deep Belief Optimization of Process Parameters for High-Quality 316L Powder Bed Fusion in Aeronautics Parts

This article addresses the issues of low shear strength and poor side surface roughness in 316L powder bed fusion caused by unoptimized key parameters. The process has been analyzed with an improved deep belief network to optimize variables such as feed rate, laser speed, power, powder preheating temperature, layer thickness, and laser radius. Constrained by actual processing limits, an selective laser melting (SLM) model is constructed to fit energy density and minimize side roughness. To solve such questions, contrastive divergence method has been applied in deep belief neural-network named CD-DBN. The adaptive hierarchical learning rate algorithm autonomously adjusts the number of hidden layers by identifying solution results to achieve better fitness, also shown in CD-DBN, which can get more accurate results while reducing the difficulty of solving. CD-DBN algorithm is trained on 12,000 training datasets and integrated with environment and printing process monitoring data; the optimal parameters are achieved and tested with nearest feasible values. A power blade has been used in verifying the result. Results indicate side roughness can be optimized to 5.34 μm and energy density can reach 16.001 by CD-DBN, which got a 17.85% improvement compared with the result calculated by other method, and the error rate of prediction of CD-DBN can be below 0.2%. This research provides a reliable way for improving the industrial application of SLM in aeronautics and astronautics fields.