Gaussian Process Regression-based Link Quality Prediction Method

Wireless sensor networks is a self-organizing network composed of notes which are deployed in the monitoring area and communicate by multi-hop.It has been widely applied in many fields.The sensor nodes communicate with low-power consumption and their environment is usually hash and complex,which leads to the instability of communication links between nodes.Perceiving the link quality information in time can provide routing reference for forwarding data,thereby effectively reducing the number of data retransmission and improving the throughput of network.Therefore effective link quality prediction mechanism is critical to improve success rate of data communication and prolong the network life.This thesis introduced the characteristics of wireless link and existing link quality prediction methods,then analyzed the definition and correlation of link quality parameters.On this basis,we proposed a novel link quality prediction model for Wireless Sensor Networks based on Gaussian Process Regression.The physical layer parameters are real-time and sensitive,but direct measurement of packet reception rate needs more energy,so we established the nonlinear mapping relationship between the physical layer parameters and packet reception rate.Since the redundancy between the link quality parameters might reduce the training speed,the thesis analyzed the grey correlation degree between them to select the effective impact factors.Then appropriate covariance functions were selected to construct the link quality prediction model based on the characteristics of link quality time series.Link communication is vulnerable to the influence and interference of space environment,geographical location,wireless signal and so on.The study object of the thesis was Wireless Sensor Networks with static nodes.We carried on the experiments to collect the experimental data from multiple pairs of nodes in different direction and distance at different scenarios,including university campus wood,laboratory in teaching building,library square and road.The thesis analyzed the link fluctuation between node pairs in different scenes and determined the input parameters of the model by analyzing grey correlation degree between different link quality parameters.In the thesis,two kinds of links were chosen to analyze the experiment and verify the model.The experimental results showed that the drop-dimensional experimental data samples still covered link quality information;The prediction performance of the model based on the combined covariance function was greater than the model's based on the single covariance function under the two links;Compared with the SVR model,the proposed link quality prediction model had better prediction accuracy.