On the Challenge of Generating Multivariate Time Series Data from Distributed Sensors in IoT-enabled Scenarios

Smart sensing environments supported by an Internet of Things (IoT) infrastructure are enabling a wide array of possibilities for data-intensive remote monitoring applications. Although nowadays it is possible to access and deploy complex setups of sophisticated sensors, the telecommunications infrastructures are struggling to deal with the massive amounts of generated data for transmission, processing, and storage from the diverse range of IoT devices. As data corruption and the loss of data are realistic scenarios in current IoT deployments, it is relevant to find an alternative solution that circumvents the limitations of the physical infrastructure for dealing with potential missing data in remote sensing environments. In this paper, we analyze the feasibility of reproducing missing sensor data due to a communication failure from correlated sources in the same experimental context. We rely on generative adversarial models for the prediction of acceleration data and evaluate the usefulness of the synthetic data with a standard human activity recognition (HAR) classifier.