Résumé: This work targets a comprehensive model enabling energy-constrained IoT (Internet of Things) sensor devices to be inactive for extended periods while estimating their readings of real-time events. Although events seem semantically uncoupled, they are usually spatially and temporally related. We propose GSAVES (Graph Sensor AdVersarial for Energy Saving), which uses readings from active devices and spatial correlations to generate the missing data due to sensor inactivity. The missing readings are generated with Graph Convolutional Network (GCN) that learns embeddings from data and the graph structure. GSAVES is evaluated against four state-of-the-art solutions using three network sizes and four performance metrics. The results demonstrate the efficiency of GSAVES for providing the best balance between the considered metrics, outperforming all the solutions in reducing energy consumption and improving accuracy.

Langue: Anglais