Deep Learning-Based Anomaly Detection for Laser-Fused Metal Components Using Pyrometric Data

Abstract Laser powder-bed fusion (LPBF) enables the fabrication of complex metal components but suffers from subsurface porosity and geometrical defects that compromise performance. Existing postprocess inspection and monitoring methods are expensive, complex, and tedious. This work presents an in situ monitoring methodology that converts high-frequency coaxial pyrometric data into three-channel images, enabling direct application of convolutional neural networks (CNNs) for defect detection. We systematically evaluate seven state-of-the-art CNN architectures (ResNet18/34/50, MobileNetV2, DenseNet121, EfficientNetB0, ShuffleNetV2) under both training-from-scratch and transfer-learning paradigms. EfficientNetB0 trained from scratch achieves the highest accuracy in our evaluation. Our results demonstrate that pyrometric imaging combined with deep learning provides a cost-effective, easily replicable approach to real-time quality assurance in LPBF processes.