Deep Learning for Process Monitoring and Defect Detection of Laser-Based Powder Bed Fusion of Polymers.

Area of Science:

• Additive Manufacturing
• Polymer Processing
• Materials Science

Background:

• Laser-based powder bed fusion of polymers (PBF-LB/P) faces challenges with part quality due to thermal instabilities.
• Defects like warping and delamination arise from crystallization within a narrow sintering window.
• Existing monitoring strategies struggle with subtle thermal deviations in industrial PBF-LB/P.

Purpose of the Study:

• To systematically evaluate diverse deep learning paradigms for defect detection in polymer PBF-LB/P.
• To assess the industrial deployment feasibility of various machine learning models.
• To establish a resource-efficient framework for real-time quality monitoring in polymer additive manufacturing.

Main Methods:

• Cross-paradigm assessment of unsupervised (autoencoders, GANs), supervised (CNNs: VGG-16, ResNet50, Xception), hybrid (CNN-LSTM), and physics-informed neural networks (PINNs).
• Utilized 76,450 synchronized thermal and RGB images from an industrial PBF-LB/P system.
• Benchmarked models on detection performance, temporal robustness, physical consistency, and computational complexity.

Main Results:

• Pre-trained CNNs achieved up to 99.09% frame-level accuracy but require significant computational resources.
• PINN models provided physically consistent thermal-field regression with an RMSE of ~27 K.
• A lightweight hybrid CNN model demonstrated 99.7% validation accuracy with minimal parameters (1860) and fast inference (1.6 ms).

Conclusions:

• Deep learning offers a scalable and efficient framework for real-time quality monitoring in polymer PBF-LB/P.
• Quantifiable trade-offs exist between model performance, complexity, and deployment requirements.
• A hybrid CNN approach balances high accuracy with computational efficiency for industrial applications.