Federated Learning for Human Activity Recognition on Resource-Constrained Sensors
IEEE Sensors Journal · 15.07.2026 · 26 (14) · 21655-21665
DOI: 10.1109/jsen.2026.3698665 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Sensors Journal, erschienen am 15.07.2026. Verfasst von Claudio N. Silva und 4 weiteren Autor:innen. Die Arbeit nennt 21 Literaturangaben. Gefördert von: Instituto Atlântico, Fortaleza, Ceará, Brazil, FAPESB INCITE.
Abstract (Originalsprache)
We present TinyFL-HAR, a lightweight and synchronous federated learning (FL) framework for on-device human activity recognition (HAR) on resource-constrained sensors. Our framework performs complete training, local evaluation, and binary model serialization directly on ESP32 microcontrollers using real-world data from the WISDM dataset, and is further validated on the PAMAP2 dataset. We conducted four experimental scenarios by varying the neural network architecture, using two fully connected variants with 2478 and 35422 trainable parameters, and the number of clients (5 versus 10), repeating each configuration across three runs. The 35k–5 Devices configuration achieved the highest accuracy (0.98), while the 2k–10 Devices configuration delivered the best recall and $F$ 1-score among the compact configurations, with training time under 100 ms. Experiments on PAMAP2 under the same four configurations provide additional evidence of robustness across datasets. In contrast, the 35k–10 Devices configuration required over 200 ms per round and substantially reduced heap memory. These findings highlight key tradeoffs in model capacity, scalability, and runtime efficiency, supporting the feasibility of FL-based HAR in highly constrained Internet of Things (IoT) environments.
Förderung
- Instituto Atlântico, Fortaleza, Ceará, Brazil
- FAPESB INCITE
Zitieren
APA
Silva, C. N., Torres, P. A., Trajano, A. F. R., Lima, J. P. B., & Prazeres, C. V. S. (2026). Federated Learning for Human Activity Recognition on Resource-Constrained Sensors. IEEE Sensors Journal, 26(14), 21655-21665. https://doi.org/10.1109/jsen.2026.3698665
Vancouver
Silva CN, Torres PA, Trajano AFR, Lima JPB, Prazeres CVS. Federated Learning for Human Activity Recognition on Resource-Constrained Sensors. IEEE Sensors Journal. 2026;26(14):21655-21665. doi:10.1109/jsen.2026.3698665
BibTeX
@article{Silva2026,
author = {Silva, Claudio N. and Torres, Pedro A. and Trajano, Alex F. R. and Lima, João Pedro B. and Prazeres, Cassio V. S.},
title = {Federated Learning for Human Activity Recognition on Resource-Constrained Sensors},
journal = {IEEE Sensors Journal},
year = {2026},
volume = {26},
number = {14},
pages = {21655--21665},
doi = {10.1109/jsen.2026.3698665}
}
EUVIMED