Semi-supervised learning for laser directed energy deposition monitoring via co-axial dynamic imaging

Data-driven approaches have emerged as a promising solution for monitoring the quality of metal parts produced by Laser Directed Energy Deposition (LDED). Current methods often require specialized knowledge for extensive data labeling, while defect characterization experiments can be costly and time-consuming. To overcome these issues, we propose a semi-supervised autoencoder framework capable of real-time prediction of local quality attributes during the LDED process. Our method significantly diverges from traditional semi-supervised learning approaches, which primarily rely on data augmentation to enforce consistency regularization. First, domain-level samples are constructed based on a series of coaxial melt pool images, providing a comprehensive assessment of the quality state of local regions by analyzing the local porosity of the samples. Second, a semi-supervised convolutional autoencoder (SCAE) is designed to develop a robust understanding of melt pool features and latent knowledge through unsupervised training. Residual connections and L2 normalization are utilized to enhance the extraction of melt pool morphological features and effectively handle the multi-scale nature of pore samples, respectively. Finally, a limited amount of labeled data is employed to activate the class space of the samples. Experimental results validate the feasibility and efficacy of the proposed model in monitoring the local quality of the LDED process. With only 60 % of the original labeled data, our model's accuracy is comparable to the maximum accuracy achieved by the fully supervised model. Even with a reduction of labeled data to 20 %, our model still maintains a prediction accuracy of 83.8 %, significantly reducing the time and costs associated with data labeling. These findings contribute to improving the quality of LDED final products and promote its broader industrial application. © 2024 Elsevier B.V.