Learning to Identify Malfunctioning Sensors in a Large-Scale Sensor Network

This paper proposes a two-stage methodology to discover malfunctioning sensors in an air quality sensor network. The two-stage methodology consists of a supervised learner to predict the future PM2.5 values of each sensor and a detector that leverages the result of the previous stage to detect the malfunctioning sensors. Consequently, even if each sensor's health status (i.e., normal or malfunctioning) is unavailable, we can still apply powerful supervised learners to this task. We conduct experiments on a nationwide air quality sensor network that includes 10,000+ sensors and utilize periodic maintenance records on some of these sensors as the ground truth of their health status. Experimental results show that this two-stage methodology can effectively discover problematic sensors. As maintaining a large-scale sensor network is laborious, the methodology can dramatically reduce the human resource required for regular inspection.