A Study of Simple Partially-Recovered Sensor Data Imputation Methods

We consider the problem of loss of continuous data feeds from sensor networks, due to transient failures. Because the failures are recoverable, part of the missing data may be, eventually, acquired. Even then, the limited resources of the nodes can result in an incomplete reconstruction of the missing data. In this paper we study a set of proposed data imputation methods, and their variations, on a real data set. We determine the tradeoffs involved in the proposed techniques. A common characteristic of the studied techniques is that they depend on the recent behavior of the data stream and do not make specific assumptions about the long-term stochastic behavior of the data. We consider also the case where simple, sub-sampling based, handling of accumulated missing data is implemented by the nodes.