Collaborative Learning at the Edge for Air Pollution Prediction

The rapid growth of connected sensing devices has resulted in enormous amounts of data being collected and processed. Air quality data collected from different monitoring stations is spatially and temporally correlated, and hence, collaborative learning can improve deep-learning (DL) model performance. Research on collaborative learning at the edge has not specifically focused so far on air quality prediction, which is the subject of this work. We compare three collaborative learning strategies and implement them on edge devices, such as the Raspberry Pi and Jetson Nano, with communication facilitated through the MQTT protocol. Federated learning (FL) is shown to enhance model accuracy in comparison to local training alone. An approach called clustered model exchange reduces communication costs during training. Finally, our proposed spatiotemporal data exchange approach exploits information from neighboring sensing stations to enhance model performance. It achieves the highest accuracy in air quality predictions, outperforming other methods in minimizing loss during training. It results in RMSE improvements ranging from 0.525% to 8.934% when compared to models that are only trained locally. We compare the real training costs of the three methods on real hardware to validate them.