Answers are grounded in this paper and cite their sources.
Federated learning successfully predicts heat emission in laser powder bed fusion, indicating improved data privacy and collaboration for SMEs.
Deep learning methods have been widely used and proved effective in defect prediction in Additive Manufacturing (AM) to ensure process stability and part quality. However, the success of deep learning models depends heavily on meticulous training, which usually requires a large homogenous dataset. This poses a challenge for the AM industry where many small- and medium-sized enterprises (SMEs) play a crucial role. On one hand, AM parts are usually customized or one-of-a-kind, and the process settings change frequently, resulting in heterogeneous datasets that usually vary by each print. On the other hand, SMEs are usually constrained by time and budget. As a result, they tend to focus on a limited number of process settings. This often leads to insufficient data collection, making SMEs difficult to properly train deep learning models independently. Therefore, there is a need to learn from the similarities in the physics of AM processes and the defect formation mechanisms, consequently enabling the potential knowledge sharing to learn across different AM scenarios. However, unique challenges in knowledge sharing arise from privacy concerns. Each design or print potentially contains sensitive proprietary information, such as process parameters, part geometries, and quality specifications. Such information may not be shared across different entities within the industry. In this context, Federated Learning (FL) emerges as a promising solution to this data scarcity and privacy challenge. FL is an innovative machine learning method that facilitates collaborative machine learning model training across multiple clients without sharing their locally stored data. In this paper, a FL framework is developed to predict section-wise heat emission, a critical process signature during the Laser Powder Bed Fusion (LPBF) for collaborative knowledge sharing across different manufacturing entities, without direct transfer of sensitive information. The framework learns the relationship between emission readings and their impact on the quality of the printed parts in LPBF across various process parameter scenarios. A Long Short-Term Memory (LSTM) model, tailored with a customized loss function, is developed for each client and trained using their local data respectively. The local models capture the time series properties under the dynamic AM process for each distinct process parameter scenario. The global model aggregates the weights from these individual local models instead of using raw data from each process directly. Three state-of-the-art FL aggregation algorithms (FedAvg, FedProx, and FedAvgM) are employed, and their effectiveness is assessed and compared through a series of experiments. Results show that the FL framework converges fast with a comparable prediction performance, when compared to training local model individually. This work demonstrates the potential of FL-enabled AM modeling and prediction where SMEs can improve their product quality without compromising data privacy.
No takes yet. Share an insight, caveat, or question.
Lei et al. (2024) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: