Xie, Y., Chen, L., Chew, Y., & Kumar, A. S. (2025). Next-Layer Defects Forecasting in Laser Directed Energy Deposition Using Convolutional Long Short-Term Memory Autoencoder. In (Editor), Volume 2B: 45th Computers and Information in Engineering Conference (CIE). https://doi.org/10.1115/detc2025-167593
Abstract:
Abstract
Laser Direct Energy Deposition (L-DED) is an additive manufacturing technology used to produce large, high-value metal components with complex geometries. However, defects such as cracks, keyhole pores, and uneven surfaces frequently occur, necessitating real-time monitoring and predictive control for quality assurance. While existing optical, acoustic, and multimodal monitoring strategies effectively identify defects, they primarily focus on defect detection rather than predictive forecasting and defect mitigation. To address this limitation, this study proposes a novel framework for next-layer defect forecasting based on historical time-series co-axial melt pool images. The key contribution is the development of a Convolutional Long Short-Term Memory (ConvLSTM) autoencoder model that captures the spatiotemporal evolution of melt pool dynamics to forecast next-layer melt pool images. The forecasted melt pool images are then fed into a Convolutional Neural Network (CNN) model to classify potential defects. The proposed hybrid ConvLSTM-CNN architecture enables proactive defect prediction by generating and classifying next-layer melt pool images. Preliminary results from single-bead wall experiments demonstrate that the ConvLSTM model achieves over 98% pixel-wise accuracy in forecasting future melt pool evolution, while the classifier independently attains over 90% defect classification accuracy. This work lays the foundation for adaptive process adjustments and preemptive defect mitigation in L-DED.
License type:
Publisher Copyright
Funding Info:
There was no specific funding for the research done