Deep Learning for Estimating Sleeping Sensor’s Values in Sustainable IoT Applications

The aim of this work is to develop a deep learning model that uses spatial correlation to enable turning turn off a subset of sensors while predicting their readings. This considerably saves the energy that would be consumed by those sensors both for sensing and communications (reporting the reading to the central station), which prolongs sensors’ lifetime and opens sky for a plethora of Internet of Things (IoT) applications. Subject of this research, event-based sensing is more challenging than periodic sensing and is uncovered in the literature. We explore advanced learning approaches including Graph Convolutional Network (GCN) and Generative Adversarial Networks (GANs) and comb them in a novel way to derive a solution that uses both spatial correlation and the readings of the active sensors to accurately generate the missing readings from inactive sensors. The proposed solution is holistic and does not rely on any duty-cycling scheduling policy. A generic random pattern is used in this paper in which every sensor is duty-cycled randomly. The structure of the network is plugged into the GCN through a graph derived using the sensing range, as well as the euclidean distance between the sensors that determines the wights on the edges. Moreover, the accuracy of the GCN is enhanced by optmizing the weights of its deep neural network with a GANs and a game theory based model, which adversarially trains the GCN’s generator by estimating the generator’s performance and calculating the Wasserstein distance between the real and the generated data. The proposed solution is evaluated in comparison with the most relevant state-of-the-art approaches in terms of accuracy, energy consumption. The results show that the proposed solution provides high performance and is clearly superior to all the compared solutions in terms of reducing energy consumption and improving accuracy.