The purpose of a wireless sensor network is to monitor, analyze, and probably respond to a time series data collected by hundreds (or even thousands) sensors distributed in the physical world. The possible applications of this network for NASA are widespread, from monitoring the ecology of an area on the planet Earth to the discovery of knowledge on Solar System Exploration. For example, to facilitate solar system exploration missions, mobile sensors mounted on robots as well as hundreds of static micro sensors can be placed on MARS to collect environment data and send them to a base station residing on MARS for real-time data analysis. The base station can then use the analysis results in real-time to determine actions that the robots should take. However, in wireless sensor networks, it can be expected that the sensor readings sent from the sensor farther into the network are lost, corrupted, or late. Sensor readings can be lost for reasons of power outage at the sensor node, sensors’ timer synchronization, random occurrences of local interferences (such as mobile radio devices, microwaves or broken line-of-sight path), a higher bit error rate of the wireless radio transmissions compared to the wire communication alternative, or a poor performance of the implemented routing algorithm in certain situations. In an effort to provide a high quality of service for a wireless sensor networks, a technique to deal with such undesirable events should be derived.

The goal of this project is to develop a data estimation technique based on association rules mining to estimate missing sensor readings. The developed technique seeks to find associations between one or more network sensors and the sensor whose data is missing, the readings of the associated sensors are then used to estimate the missing sensor value. This technique is particularly suited for data stream environments where data volume grows continuously and possibly unboundedly. To accomplish this, the technique is made computationally aware of the need to minimize both the time required to generate association rules as well as the physical space needed to store the data. This technique achieves good quality of service (QoS) for NASA applications running in sensor networks, where QoS is defined as a function of the correctness of the estimated data and the achieved response time for producing the estimation, while paying close attention to the power consumption by the sensors. Currently, we are researching a combined temporal and spatial data mining method to use on top of the association rule mining technique. We anticipate that incorporating knowledge on temporal and spatial aspects of sensor data into the data estimation technique can help improve estimation accuracy, energy consumption, and execution time and, hence, further assuring a good QoS.

A simulation version of this project has been implemented and tested on two test applications.

Click here for bios and contact information of the Data Stream research team.

Click here to link to the Data Stream publication page.

Click here to watch the software DEMO for the paper "Towards Accurate Estimation of Sensor Data Stream: A Data Mining Approach" the video DEMO page.