Dealing with Dynamic-Scale of Events: Matrix Recovery Based Compressive Data Gathering for Sensor Networks

摘要

The mass data produced in sensor networks has triggered a large variety of applications, e.g., smart city, environmental monitoring, etc. However, gathering such data from vast number of sensors throughout the network is a daunting and costly work. Previous works suffer from either a high communication overhead or a poor data recovery resulted in compressive sensing due to the high risk of sparsity violation. This paper introduces a new data gathering method to address two problems: one is how to compress and gather the large volume data effectively, the other is how to keep various time/space-scale event readings unaltered in the gathered data. According to state-of-the-art, either problem can be solved well but never both at the same time. This paper presents the first attempt to tackle with both problems simultaneously for sensor networks, from theoretical design to practical experiments with real data. Specifically, we take advantage of the redundancy and correlation of the sensor data cross time and spatial domain, and based on which we further introduce our low-rank matrix recovery design effectively recovering the gathered data. The experiment results with real sensor datasets indicate that the proposed method could recover the original data with event readings almost unaltered, and generally achieve SNR 10 times (10db) better than typical compressive sensing method, while keeping the communication overhead as low as compressing sensing based data gathering method.