Fuzzy Support Vector Regression-based Link Quality Prediction Model For Wireless Sensor Networks

Wireless sensor network is a self-organizing multi-hop network, including a large number of low power nodes deployed in monitoring area. Link is the key element to achieve interconnects and multi-hop communication. Link quality is the fundamental of upper protocols, such as topology control, routing, and mobile management. The effective link quality prediction(LQP) can improve network throughput, and can prolong network life time.In this thesis, a brief view is presented on the related works on WSN link quality parameter and link characteristics, and classification and comparison of link quality prediction model are carried out. Considering the disadvantages of current method, we present a novel link quality prediction model, chaos particle swarm optimization and fuzzy support vector regression(CPSO-FSVR), which combines the information from link layer and physical layer to describe link quality better, namely, received signal indicator(RSSI), link quality index(LQI) and signal noise rate(SNR), packet received rate(PRR). Our model introduces the principal component analysis(PCA) to transfer the feature, so as to reduce data. Taking the impact of unstable links in communication into consideration, a kernel fuzzy c-means(KFCM) algorithm, an unsupervised learning algorithm, is applied to clustering the training set automatically. As the impact of noise and outliers which generated by interference and environment can not be ignored, the membership degree of samples obtained from KFCM is to get fuzzy set for FSVR to get high accuracy. Finally, the chaos particle swarm optimization(CPSO) algorithm having good ergodicity and convergence is employed on each cluster in order to choose the suitable parameter combination for the link quality prediction model.Because link quality dependents on many factors, i.e., distance between nodes,multi-path propagation and interference from adjacent channels, we collect the link quality samples from different scenarios, including both indoor and outdoor(square,playground and road). We show that RSSI has correlation with distance, while the packet received rate has large variations at a certain distance, and multi-path propagation and interference has significant mpact on communication. PCA and KFCM are introduced to preprocessing the samples. Experiment results show that reduction can maintain accuracy.In comparison with two parameters optimization algorithm:grid and genetic algorithm, the CPSO are more efficient in their perception of link change both in stable and unstable link. The experiment results verify that compared with empirical risk-based logistic prediction methods, CPSO-FSVR model with less probe package can achieves higher accuracy and better generalization ability.