Personalized federated unsupervised learning for nozzle condition monitoring using vibration sensors in additive manufacturing

Abstract

Additive manufacturing (AM), particularly the fused filament fabrication (FFF) process, enables the production of personalized products with unique features. However, the FFF process is prone to issues such as nozzle clogging, which can degrade print quality or cause print failure. Data-driven approaches present viable solutions for real-time monitoring and defect identification in AM, enhancing both the precision and reliability of the FFF process. Despite these advantages, practical deployment faces obstacles including limited availability of high-quality data, significant labeling costs, and the rarity of anomalous data. While similar data may exist across other AM manufacturers or machines, data centralization and sharing are often constrained by privacy and competition concerns. This paper introduces FULAM, a personalized federated unsupervised learning method designed to detect anomalies in FFF machine vibration data. The framework addresses critical challenges such as data privacy, heterogeneity, and labeling costs by enabling collaborative training of unsupervised anomaly detection models across multiple clients while keeping data decentralized. A systematic analysis and comparison of recent unsupervised deep anomaly detection methods of varying complexity, traditionally evaluated in centralized settings, is conducted under federated learning (FL) environments to identify the most effective model for FFF machine vibration data. Experimental results highlight the personalized adaptation and regularization benefits of FULAM, showing cases where it outperforms both centralized approaches and state-of-the-art FL algorithms. FULAM demonstrates potential for developing robust anomaly detection models, advancing real-time condition monitoring in AM.