| The wireless sensor networks monitors the sensing area by adapting a large number of sensor nodes.The node can't carry too much energy and always deployed in some harsh environments.The communication process is easy to be interfered,resulting in instable link,and there will lose some packets during the transmission process.If the link quality can be predicted before the data packet by the node sends,and the node selects the link which the quality is good for data transmission,that can reduce the retransferring of the data packet and save the limited energy of the node.Hence,it can improve the life cycle of the entire network.The thesis analyzes the link characteristics of wireless sensor networks,and introduces the existing link quality prediction methods,selecting the mean,variance and asymmetry of the wireless link's physical parameters as the link quality parameter.Zero-phase component analysis(ZCA)whitening is applied to remove the correlation between sample to preprocess the sample.The Gaussian mixture model algorithm based on unsupervised clustering is used to divide the link quality level,it can avoid the error interference caused by the artificially selected critical point on division the level.For the imbalance problem of the wireless link sample,K-means Synthetic Minority Oversampling Technique(SMOTE)is used to process the unbalanced sample,and the Kmeans clustering is used to separate the sample into different clusters,the number of samples of a few classes is increased by random linear interpolation in each cluster.Therefore,the number of samples in every level can reach to a new balance.The Link Quality Estimator Based on The Weighted Random Forest Classification(LQE-WRFC).The decision tree with low classification performance is assigned less weight,and the decision tree with higher classification performance is assigned more weight.Constructing the time series sample set by sliding window,Link Quality Prediction Model Based on Random Forest Regression(LQP-RFR)to predict the link quality level at the next moment.The experimental results in five scenarios show that the estimation model has better performance which processed by K-means SMOTE.LQE-WRFC can obtain better estimation performance than the Naive Bayesian,Logistic Regression and K-Nearest Neighbor estimator.Compared with Exponentially Weighted Moving Average,Triangle Metric,Support Vector Regression and Linear Regression prediction models,LQP-RFR has higher prediction accuracy. |
