A Deep Learning Approach to Sensor Fusion Inference at the Edge.

The advent of large scale urban sensor networks has enabled a paradigm shift of how we collect and interpret data. By equipping these sensor nodes with emerging low-power hardware accelerators, they become powerful edge devices, capable of locally inferring latent features and trends from their fused multivariate data. Unfortunately, traditional inference techniques are not well suited for operation in edge devices, or simply fail to capture many statistical aspects of these low-cost sensors. As a result, these methods struggle to accurately model nonlinear events. In this work, we propose a deep learning methodology that is able to infer unseen data by learning complex trends and the distribution of the fused time-series inputs. This novel hybrid architecture combines a multivariate Long Short-Term Memory (LSTM) branch and two convolutional branches to extract time-series trends as well as short-term features. By normalizing each input vector, we are able to magnify features and better distinguish trends between series. As a demonstration of the broad applicability of this technique, we use data from a currently deployed pollution monitoring network of low-cost sensors to infer hourly ozone concentrations at the device level. Results indicate that our technique greatly outperforms traditional linear regression techniques by 6x as well as state-of-the-art multivariate time-series techniques by 1.4x in mean squared error. Remarkably, we also show that inferred quantities can achieve lower variability than the primary sensors which produce the input data.