基于门控循环单元的链路质量预测

Link Quality Prediction Based on Gate Recurrent Unit

无线传感器网络中,节点传输数据时容易受到环境中噪声的干扰,使传输链路质量变差,导致数据包丢失、消息重发,从而加速节点能量的消耗,缩短网络寿命。链路质量预测可以为上层路由协议选择高质量的无线链路进行通信提供依据,通过链路质量预测选取高质量的链路传输数据,可以提高数据传输效率,减少重传次数。本文提出基于门控循环单元(gate recurrent unit,GRU)的链路质量预测方法。首先,采用Gap Statistic算法优化的Kmeans++算法,依据包接收率的分布自适应划分链路质量等级,获得链路质量样本标签;其次,选择接收信号强度均值、链路质量指示均值及信噪比均值作为输入,借助CatBoost在分类问题上的优势,构建链路质量评估模型,并采用网格搜索法对参数寻优;最后,使用滑动时间窗口构建链路质量等级时序样本集,使用GRU提取链路质量等级时间序列的时序信息,为进一步提高预测的准确率,采用支持向量回归机构建链路质量预测模型,预测下一时刻链路质量等级。本文采用真实场景中的数据进行实验,根据主要干扰源不同,选择实验室、走廊和停车场3个场景收集数据,使用均方误差评价链路质量预测模型的有效性。实验结果表明,与小波神经网络、循环神经网络和随机向量函数链等方法相比,所提方法具有更小的预测误差,可以准确预测链路质量等级。

In wireless sensor networks,nodes are susceptible to noise interference,which makes the transmission link quality worse,leading to packet loss and retransmission,thus speeding up the energy consumption of nodes and reducing the network life.Link quality prediction can provide a basis for upper layer routing protocols to select high-quality wireless links for communication,which can improve data transmission efficiency and reduce the number of retransmissions.A link quality prediction method based on gate current unit(GRU)was proposed in this paper.Specifically,the K-means++algorithm optimized by the gap statistic algorithm was used to adaptively divide link quality levels according to the packet reception rate.In this way,the labels of link quality samples were obtained.Then,the received signal strength indicator mean,the link quality indicator mean and the signal to noise ratio mean were selected as the input.A link quality estimation model was constructed based on CatBoost because of its strength in classification,and the grid search optimization algorithm was employed to optimize the parameters of estimation model.Finally,the link quality level time series sample was constructed by sliding windows according to the results of the estimation model.GRU was used to extract link quality time series information.In order to improve the accuracy of prediction,the timing information was input to support vector regression to predict the link quality level at the next moment. Data collected from the real world were used in this paper. Accord-ing to the different common interference types, three scenes of lab, corridor and parking lot were taken as experiments scenarios. The Mean square error was used to evaluate the performance of link quality prediction models, and the experimental results showed that compared with wavelet neural network, recurrent neural network, and random vector functional link and so on, the proposed method can predict the link quality level ac-curately.