Extracting Critical Attributes in Wireless Sensor Networks Using Modified Regression Polynomial

Abstract

In this paper, the spatio-temporal nature of sensor data is exploited to approximate current values of the sensors based on the observation of the same sensor in the previous period of time and the neighboring sensors’ data. By applying linear regression, a polynomial can fit variation in the data over a given region for a specific period of time. Assuming attributes to have a gradual change both over time and space, readings of the neighboring sensors are highly correlated and the readings of sensors at any current period of time do not change drastically from the previous period. Based on this assumption, instead of performing the regression algorithm every time the sensed attributes change, an iterative scheme is employed by which the current values of sensors are approximated based on the previous readings. The percentage error in this approximation process is found to be less than 6%, while saving considerable amount of computational overhead. Analysis of the regression polynomial also accurately gives the location of critical points in the sensed region where minimum and maximum attribute values occur and could be very useful for sensor network applications.