Speaker
Description
Laser Powder Bed Fusion (L-PBF) of aluminum alloys is highly sensitive to local thermal conditions, where subtle variations in melt pool dynamics can lead to defect formation such as lack of fusion, porosity, or gas trapping. Detecting and characterizing such defects during the build process remains a key challenge for in-situ monitoring systems. In this work, an infrared thermography (IRT)–based monitoring framework combined with the extraction of suitable thermal features and physics-informed machine learning is presented for the detection and characterization of imposed typical Additive Manufacturing (AM) defects intentionally introduced in the CAD design of specimens produced via Laser Powder Bed Fusion process (L-PBF). The aim of this work is to identify and characterize these defects with an online non-destructive evaluation approach, without, without requiring a post-process inspection.
L-PBF experiments were monitored on AlSi10Mg powder using an off-axis microbolometer infrared camera to capture thermal image sequences during multi-layer fabrication. The thermal data were first processed to remove the countering scan strategy and to extract the spatial footprint of the printed cube using maximum-temperature projections. For each layer, a binary printing mask was generated to exclude background and powder spreading effects. Pixel-wise temporal temperature signals within the printed region were then analyzed.
To handle scan-order effects, each pixel’s thermal signal was aligned to its local heating peak, converting absolute time into an event-based reference frame. From the aligned signals, physics-informed thermal features describing the local heating and cooling behavior were extracted, including peak temperature, heating rate, early-time cooling slope, exponential cooling time constant, goodness-of-fit of the cooling model, and dwell time above a thermal threshold. These features encode melt pool stability, heat dissipation efficiency, and thermal consistency of the process parameters closely linked to defect formation mechanisms in L-PBF.
As a preliminary approach and with the aim of an automatic defect identification, unsupervised defect identification was performed using a Gaussian Mixture Model (GMM). Initially, clustering was applied on a per-layer basis to identify distinct thermal behavior regimes within individual layers. To enhance sensitivity to imposed defects and avoiding phenomena related to multi-pass strategy and edge effects, a multi-layer aggregation strategy was then introduced, where pixel-wise features were combined across several consecutive layers to form thermal history descriptors for each spatial location. Clustering these multi-layer descriptors enabled the identification of persistent anomalous regions that correlate with the imposed defect geometries defined in the CAD model.
The extraction of physics-informed thermal features, together with the resulting feature-based cluster maps, revealed spatially coherent regions aligned with the imposed defects, exhibiting abnormal cooling behavior and poor exponential fit quality that were not apparent in single-layer analyses. These regions aligned with the known locations of imposed defects, demonstrating the capability of the proposed approach to detect defect signatures for an on-line non-destructive evaluation. Further analyses were than carried out to extract information regards defect characterization, considering different defect geometries and type of typical AM defects.
The study demonstrates the effectiveness of infrared thermography combined with physics-informed thermal feature analysis and a machine learning approach for in-situ L-PBF process monitoring and defect identification and characterization, paving the way toward intelligent, data-driven quality assurance in additive manufacturing.
