[Data] Optimizing In-situ Monitoring for Laser Powder Bed Fusion Process: Deciphering Acoustic Emission and Sensor Sensitivity with Explainable Machine Learning

Intentional variations in Metal-Laser Powder Bed Process (L-PBF) spaces were introduced by employing two distinct distributions of 316L stainless steel powder, characterized by particle sizes both greater than 45 μm and less than 45 μm. These powders underwent processing with two different sets of laser parameters, resulting in the generation of four datasets (D1, D2, D3, D4). These datasets encapsulate airborne Acoustic Emissions (AE) arising from diverse build qualities, including Lack of Fusion (LoF) pores, conduction mode, and keyhole formations. The experiments were conducted utilizing a Sisma MYSINT 100 commercial LPBF printer and an airborne Acoustic Emission (AE) sensor system boasting a flat frequency response ranging from 0 to 150 kHz. Validation of the ground truths for the three laser regimes across the four datasets, representing distinct process spaces, was accomplished through the confirmation of cross-sectional images. In the fabrication of a cube using a powder bed and laser, data acquisition from an AE sensor commenced when the optical intensity reached a threshold of 0.5 V for each scan length. The photodiode trigger gain was adjusted to saturate at 5 V, and the subsequent continuous-time window, where the optical signal remained at 5 V for 12.5 ms, was calculated and segmented to generate the dataset. Irrespective of the specific regime (Lack of Fusion, Conduction, and Keyhole) or the cube being fabricated (with two powder distributions), the signals obtained during this process were segmented into a 12.5 ms window comprising 5000 data points. Each dataset is accompanied by two files – one for raw data and the other for ground truth labels. To mitigate noise, an offline low-pass Butterworth filter with a 150 kHz cut-off frequency was applied, aligning with the frequency response specification of the AE sensor. The total number of AE windows per dataset is approximately 7500 raw AE signals.