Link Quality Estimation Method Based On Gradient Boosting Decision Tree

In the application of wireless sensor networks,link quality estimation is the basic problem to guarantee the reliable transmission of data and the performance of the upper layer network protocol.Effective link quality estimation methods can not only objectively evaluate link quality,link stability and agility,but also provide guarantee for data transmission.This thesis comes from the national natural science foundation of China and aims to study the link quality estimation method in wireless sensor networks.As a machine learning method,Gradient Boosting Decision Tree(GBDT)has strong robustness and high precision.In this thesis,link quality estimation is transformed into a multi-classification problem,and the nonlinear correlation between physical layer parameters and packet received rate is analyzed by using the maximum information coefficient method.The received signal strength indication,link quality indicator and signal noise rate are selected as the input parameters of the link quality estimator.Considering the influence between outliers and different dimensionality of parameters,the boxplot method is used to screen outliers and carry out smoothing and normalization processing to reduce the complexity of the model.The link quality grade is divided according to the packet received rate and used as the link quality estimation index.A link quality estimator based on GBDT is established by choosing the minimization negative logarithmic loss function.To solve the problem that parameter setting has a great influence on the estimation result,an improved particle swarm optimization algorithm with mutation operator and inertia weight is adopted to optimize the parameters of GBDT estimator.In this thesis,wireless sensor network is deployed in indoor and outdoor environment(indoor corridor,campus forest and campus road)and uses a link quality test platform software developed by our laboratory to obtain experimental data.Accuracy rate,precision rate,recall rate and F1 index are adopted to evaluate the performance of GBDT estimator and other estimators.Experimental results show that the GBDT estimator optimized in this thesis has the best performance.Compared with traditional GBDT estimator and Support Vector Machine(SVM)estimator,this estimator has higher accuracy and stability.