Publikation
Federated Multi-Modal Learning for Manufacturing: A Privacy-Preserving Approach to Distributed Sensor Fusion
Tatjana Legler; Vinit Vikas Hegiste; Martin Ruskowski
In: 2025 3rd International Conference on Federated Learning Technologies and Applications (FLTA). International Conference on Federated Learning Technologies and Applications (FLTA-2025), Pages 562-570, IEEE, 2025.
Zusammenfassung
The integration of multi-modal data sources in manufacturing environments presents significant opportunities for enhanced process monitoring, quality control, and predictive maintenance. However, traditional centralized machine learning approaches face challenges in industrial settings due to data privacy concerns, network limitations, and the distributed nature of manufacturing operations. This paper explores the applica tion of federated learning to multi-modal manufacturing data, examining how distributed learning frameworks can leverage heterogeneous data while preserving data locality. We implement and evaluate a modular federated multi-modal learning frame work capable of handling heterogeneous modality availability across clients through dedicated feature extractors and flexible fusion strategies. The approach supports partial participation, enabling sites with only a subset of modalities to contribute to the global model without imputation. Our results demon strate that federated multi-modal learning maintains competitive accuracy compared to centralized approaches while preserving data privacy. Notably, we find that distributed missing modal ities can actually enhance system performance as clients with heterogeneous modality availability outperform traditional IID federated setups by up to 8 percentage points through com plementary knowledge aggregation. The framework effectively handles missing modalities, demonstrating that their absence can be viewed as an opportunity for specialization rather than as a limitation. These findings highlight the potential of federated multi-modal learning for realistic manufacturing scenarios with diverse sensor configurations and privacy constraints, where different sites naturally possess varying sensor capabilities.