| Nodes in wireless sensor networks are limited by energy and transmission distance.Data is sent to the destination node by forwarding.Nodes are susceptible to noise in the environment when transmitting data,so that the link quality is poor,leading to packet loss,message retransmission,speeding up consumption of energy,and shortening life span of networks.Selecting high-quality link to transmit data with the help of link quality prediction mechanisms can reduce the number of retransmissions and extend the life of the network.After analyzed the link characteristics of wireless sensor networks,the received signal strength mean,link quality indication mean and signal-to-noise ratio mean were selected as the link quality parameters.K-means++ algorithm optimized by gap statistic was employed to grade link quality level,in terms of which link quality is determined.Considering the missing data in the link quality samples,the mean filling method was taken to fill the missing ones.Because of training a link quality estimation model with unbalanced samples would bias the results of the model towards the majority class of samples,the SMOTETomek algorithm was adopted to deal with the link quality sample imbalance problem.Link quality estimation model was constructed based on Cat Boost due to its advantage of classification.Grid search optimization algorithm was employed to optimize super parameters of estimation model,such as learning rate,depth of tree,and etc.Temporal sample set of link quality level was constructed by sliding window according to the results of estimation model.Link quality prediction model based on GRU was constructed,and the temporal feature in sample set was learned by GRU.With input of temporal feature,support vector regression was employed to predict the link quality level at the next moment.Link quality parameter samples were collected in three scenarios with different interference,such as lab,corridor and parking lot.The experimental results showed that the effect of estimation was improved after processing sample imbalance.Compared with link quality estimation models constructed by support vector machine,random forest,fuzzy rules and other methods,the proposed model had higher accuracy,precision rate and F1 score.Compared with the link quality prediction model based on wavelet neural networks and recurrent neural networks,the proposed link quality prediction model had smaller prediction error. |
