I-WSN League: Clustered Distributed Learning in Wireless Sensor Networks

Abstract

In this work, i-WSN League, a comprehensive hardware/software framework for the support of distributed training and inference is introduced. For what concerns the hardware, in i-WSN League two types of nodes are considered, namely, head nodes and common nodes. Head nodes are resource-rich nodes that have the capabilities for training artificial neural network. Common nodes collect data and can execute inference only. In i-WSN League, all nodes are grouped in Clusters, each with a Cluster Head (selected among the head nodes), which is the only node responsible for training. To this end, the data coming from all nodes in the Cluster can be utilized. This, however, involves a large exchange of data which might be unsustainable by common nodes. Thus, only part of the data collected by common nodes is sent to the Cluster Heads and a network of Cluster Heads will implement distributed learning in a peer-to-peer fashion. As compared to state-of-the-art literature, the key contributions of our work are related to the combination of gossiping and clustering to adapt the operations executed by each node to its capabilities, with the aim of minimizing the energy consumption in resource-limited nodes, while preserving accuracy. In this article, i-WSN League is assessed in a simple scenario in which a wireless sensor network monitors the air pollution in a large city. Performance results obtained by considering auto-encoders prove the effectiveness of the proposed scheme as well as its balanced energy consumption and fairness in resource consumption distribution.