Federated learning-based semantic segmentation for pixel-wise defect detection in additive manufacturing

Abstract

Semantic segmentation is a promising machine learning (ML) method for highly precise fine-scale defect detection and part qualification in additive manufacturing (AM). Most existing segmentation methods utilize convolutional neural network architectures that require large quantities of training data. However, obtaining sufficient data—both in quality and quantity—to train such models is expensive and time-consuming for individual AM practitioners, which severely limits the deployment of semantic segmentation in a data-scarce production environment. Similar data may be readily available with other AM practitioners that cannot be pooled together for conventional centralized learning (CL) due to its sensitive nature or conflicts of interest. This paper develops a federated learning (FL)-based method to simultaneously alleviate the constraints of data availability and data privacy. A U-Net architecture is created for semantic segmentation and is trained under the FL framework. The effectiveness of the developed FL-based semantic segmentation approach is demonstrated using case studies on layer-wise images from the laser powder bed fusion process. Results show that the proposed technique achieves a comparable defect detection performance with CL, which shares data among manufacturers/clients but does not preserve data privacy, and significantly outperforms individual learning, where each manufacturer trains a model using its own data. Additionally, the impact of data distribution across clients, incentives to participate in FL, and the learning dynamics of FL are discussed in detail. It is found that data diversity within and across clients improves FL performance, and FL does not involve a significantly higher training cost compared to CL. Lastly, transfer learning is shown to enhance FL generalizability, thus allowing more manufacturers with heterogeneous machines or technologies to benefit from participating in a data federation. Overall, this work puts forth FL as a promising paradigm for privacy-preserving collaborative ML in AM process control.