Comparing LR, GP, BPN, RBF and SVR for Self-Learning Pattern Matching in WSN Indoor Localization

It is a challenging issue to apply WSN (Wireless Sensor Network) to achieve accurate location information. PM (Pattern Matching), known as one of the most famous localization methods, has the drawback of requiring high initialization effort to predict/train MF (Matching Function). In this paper, the authors propose SPM (Self-learning PM) to improve not only the localization accuracy but also the initialization effort of PM. SPM applies a divide-and-conquer self-learning scheme to reduce the number of training patterns in training. Additionally, it introduces a Bayesian filtering scheme to remove the noise signal caused by multipath effects so as to enhance localization accuracy accordingly. This paper applies different training methods (linear regression, Gaussian process, backpropagation network, radial basis function, and support vector regression) to evaluate the performances of SPM and PM in a complicated indoor environment. Experiments show that SPM is better than PM for all training methods applied. SPM can use up to 72% fewer training patterns than PM to achieve the same localization accuracy. If the same number of training patterns is utilized, SPM can achieve up to 58% higher localization accuracy than PM.