Hybrid AI Boosts Real-Time Melt Pool Monitoring in 3D Printing

Real-time monitoring of the melt pool is critical for quality control in laser powder bed fusion (LPBF) additive manufacturing. Defects in the melt pool can lead to flawed parts, making rapid anomaly detection essential. This research explores the application of AI and machine learning for this purpose, focusing on binary image classification to distinguish normal from abnormal melt pool images. The study benchmarked several transfer learning architectures (ResNet50, EfficientNetB0, MobileNetV2) against two Random Forest approaches: one using raw pixel features and another, hybrid model, trained on features extracted from EfficientNetB0. The goal was to find a solution that balances high accuracy with low inference time, suitable for the hardware constraints of open-architecture LPBF machines. The hybrid EfficientNetB0-plus-Random Forest model emerged as the top performer. It achieved an F1 score of 0.9451 and an accuracy of 0.9458, with an impressive sub-millisecond inference time of 1.15 ms per image. This demonstrates that combining powerful pre-trained convolutional features with classical ensemble methods offers a robust and computationally efficient solution for real-time anomaly detection in data-limited manufacturing environments, significantly outperforming purely deep learning models in inference speed while maintaining accuracy.