A deep learning framework for defect prediction based on thermographic in-situ monitoring in laser powder bed fusion

Abstract The prediction of porosity is a crucial task for metal based additive manufacturing techniques such as laser powder bed fusion. Short wave infrared thermography an in-situ monitoring tool enables the measurement surface radiosity during exposure. Based on thermogram data, thermal history component can be reconstructed which closely related to resulting mechanical properties and formation in part. In this study, we present novel framework local extracted features from data. consists data pre-processing workflow supervised deep learning classifier architecture. generates samples feature by including information multiple subsequent layers. Thereby, occurrence complex process phenomena keyhole pores enabled. A custom convolutional neural network model used classification. trained tested dataset thermographic AISI 316L stainless steel test component. impact parameters void distribution classification performance studied detail. presented achieves accuracy 0.96 f1-Score 0.86 predicting small sub-volumes with dimension (700 × 700 50) µm 3 . Furthermore, show that threshold sample labeling number included layers are influential performance. Moreover, shown sensitive changes although it binary labeled disregards actual porosity.