BLESS: A Divine Energy Saving Protocol for Wireless Sensor Networks
Smart cities use real-time data to improve the management of cities' resources. Wireless sensor networks (WSNs) specially help in collecting this data, such as pollution, traffic etc. On the other hand, the energy resources of sensor nodes are constrained and some operations such as sensing and wireless communication consume the most energy [1]. The energy consumption can be reduced by estimating the sensor readings of some sensor nodes and putting them to sleep. This can be accomplished by training prediction models to estimate the sensor readings. WSNs are usually in a dynamic environment, so the prediction models can become stale. Therefore these stale models need to be identified energy-efficiently.
Our approach is focused on saving energy in WSNs by scheduling some sensor nodes to sleep, while estimating the readings of these nodes using prediction models. In particular, we propose an energy-efficient approach to identify the goodness of the prediction models, based on our trust in the model.
State of the Art
In Dual Prediction Schemes [2] (DPS), the trained prediction model resides both in the sensor node and in the central base station. Although only wrong estimates of the readings are communicated, sensing needs to be done by all the nodes all the time. The most importand issues with these schemes are the energy drain due to always-on sensing operations and the fact that they are not suited to large scale sensor networks.
In order to provide a solutions to DPS issues, a set of active and sleeping nodes are identified; the active nodes' readings are used to estimate the values of the sleeping nodes [3]. So far, periodic monitoring of sleeping nodes is the only approach to detect stale prediction models. New efficient monitoring techniques are needed to improve the trade-off between the time required to identify when models are stale vs the energy consumed.
Trust in prediction models
We propose to use simple linear regression (SLR) as the prediction model
Yt = A + Bxt + et
Autocorrelation exists in the error term
- Trends are identified based on the autocorrelation to estimate future errors.
- If future errors exceed a user-specified threshold (η), our trust in the model is low.
The trust of the model is given either as a positive or negative feedback based on how well it can estimate the future errors. The sleeping node's monitoring frequency can be adapted according to the trust in the model.
Our approach
Results
The results shown are based on the temperature sensors of the Intel Lab [4] for 10 days. We outperform Koushanfar et. al [3] in terms of the relative update rate because of the expensive training and retraining of their prediction model. [3] does not exhibit any errors because of the careful pre-planning of the monitoring for the given dataset, which cannot be generalised to all WSN deployments.
People
Mithileash Mohan, Mélanie Bouroche, Vinny Cahill
Sponsors
