Abstract:

In order to improve the accuracy of the general localization method for wireless sensor network （WSN） due to the big inter-node ranging error contributed by the coarse ranging technology, a new localization method based on the least-square support vector regression （LS-SVR） of feature importance is proposed. In this method, the range from the unknown node to the anchor node is taken as the feature variable, and the feature is extracted according to the importance of the feature variable. Training samples are obtained via the gridding in the detection region and are then studied via the LS-SVR to establish a localization model. In the localization phase, the feature vector of the unknown node is input into the localization model, and the accurate location of the unknown node is achieved by utilizing the good generalization capability of LS-SVR. A localization experiment is finally performed with 100 uniformly-distributed nodes and 100 randomly-distributed nodes in the C-shape region. The results show that the proposed method effectively eliminates the influence of ranging error on the localization accuracy and it reduces the average location error to 7.5% - 14. 0% less than that of the DV-Hop method in uniform distribution condition and 36.5% -55.2% less in random distribution condition in the C-shape region.

Key words: feature extraction, least-square support vector regression machine, wireless sensor network, localization