A Fuzzy Logic based Offloading System for Distributed Deep Learning in Wireless Sensor Networks

In recent years, wireless sensor networks have been deployed on a large scale. Their ability to capture information and transmit it has made them key players in the smart city and industry. Most often, data streamed from sensor nodes are collected in a centralized cloud to form a vector of features. They are then used to make future predictions using advanced machine learning (ML) techniques such as Deep Learning (DL). Due to the increasing volumes of data being collected and transmitted over the network, this approach turns out to have a high communication cost, which wastes energy and often saturates the network. In this work, we study an alternative approach that enables collaborative inference throughout the network devices (edge, fog and cloud). Our proposition consists of a fuzzy logic controller which will decide at what level of the network (edge, fog or cloud) the inference should be carried out. Three parameters are taken into account to make this decision: the available energy level of the edge device (smart node), the available network bandwidth and the amount of available data for inference. This work presents preliminary results achieved with the proposed approach using generated configuration input data. We show that the implementation of our controller on edge devices (smart nodes) that run RNN-LSTM for multivariate time series predictions can reduce their energetic cost by around 50%.