In-Situ Monitoring and Quality Assurance Algorithms for Laser Powder-Bed Fusion using Optical Tomography

Abstract

Conventional methods for evaluating the quality of components produced through laser powder-bed fusion additive manufacturing (LPBF) are often costly and resource-intensive. These post-production techniques involve mechanical testing, detailed metallographic examination, and non-destructive methods like X-ray computed tomography (CT) to identify flaws. Recently, there has been a shift towards employing in-situ monitoring systems, such as optical tomography (OT), which capture near-infrared light emissions to detect defects arising during LPBF. This dissertation introduces innovative approaches for defect detection in LPBF, utilizing OT data alongside machine learning techniques. LPBF processes inherently exhibit random behavior, presenting challenges in developing robust defect detection algorithms adaptable to diverse machine setups and process parameters. The proposed model integrates a self-organizing map (SOM), a fuzzy logic scheme, and a tailored U-Net architecture to detect and predict defect probabilities in LPBF-produced parts using in-situ OT analysis. Specifically, the model is designed to identify common flaws such as lack of fusion and keyhole defects.

The effectiveness of the approach was validated through a series of experiments. Initially, the influence of process parameter selection on recorded in-situ optical tomography (OT) data was investigated. This was followed by the intentional and random recreation of process defects to simulate the stochastic nature of real-world manufacturing processes and to gain a deeper understanding of defect formation. The developed model was subsequently evaluated on a complex geometry to assess its performance in a practical setting. Validation of the model was done by comparing its predictions against computed tomography (CT) scans, to achieve this, Dynamic Time Warping (DTW) technique was used to measure the similarity between porosity curves generated by the model and those from CT scans. The developed model effectively predicted porosity resulting from lack of fusion or keyhole defects across various process parameter settings, achieving average Euclidean distance scores of 0.243 for lack of fusion pores and 0.6 for keyhole pores. Additionally, the model successfully detected defects in complex geometries with internal lattice structures.

A significant advantage of this developed model is its adaptability. Fuzzy logic allows for the integration of soft decision boundaries and expert rules into the model, which is crucial when dealing with complex phenomena like porosity where the boundaries between the presence of defects in the fabricated part, based on monitoring OT data, are not always clear-cut. Expert knowledge can be encoded into fuzzy rules that mimic human reasoning and decision-making processes. Quality assurance experts can use their experience to provide insights through the application of fuzzy rules, determining how certain visual or measurable features of an image typically correspond to specific types of porosity. They can also adjust the probability threshold for defect detection based on specific quality criteria. This adaptability enhances the approach's utility across diverse manufacturing scenarios, offering flexibility in meeting quality assurance requirements.