Data imputation on IoT gateways using machine learning

IoT (Internet of Things) gateways receive data from thousands of sensors and send it to the cloud, which runs intelligent services. However, collected data might have missing or anomalous values due to various reasons, such as network problems, damaged sensors, or security attacks. Missing and noisy data can affect future decision-making, so IoT gateways need to transmit consistent data to the cloud. This work proposes a method to impute missing data on IoT gateways based on neural network regression. We validate this method using six years of weather data from a station located in Rio de Janeiro, considering different percentages of missing data. The results show that the regression models have more than a 0.92 R-squared score and low errors when predicting sensor measurements. Furthermore, we show that the neural network implementation can run on IoT gateways due to its short execution time and low memory utilization. Finally, we show that a single model performs well even when 50% of the data is missing, highlighting the proposed approach's generality.