Research On Spectrum Prediction Techniques In Cognitive Vehicular Networks

As an indispensable part of the intelligent transportation system,the vehicular network has become an important means to ensure road traffic safety,improve traffic management ef-ficiency,and optimize transportation experience through the deployment of sensors and the integrated application of wireless communication technology.However,with the rapid devel-opment of the global automobile industry,high-density traffic flow and road congestion have become very common.A large number of gathered vehicles frequently compete for limited communication resources,causing serious delays and packet loss.In this case,it is difficult for the existing communication technologies to meet the requirements of various safety applications and non-safety applications in the vehicular networks by ensuring reliable and efficient infor-mation transmission.Different from the situation of high communication demand and shortage of communication resources in the vehicular networks,some licensed channels and unlicensed channels that are fixedly allocated to other application scenarios are faced with problems such as low efficiency and waste of spectrum resources.Therefore,introducing cognitive radio tech-nology into the vehicular network to build a cognitive vehicular network,allowing vehicles to obtain other licensed channel resources opportunistically,has become an effective means to alleviate the contradiction between the high spectrum requirement and the spectrum shortage problem.At present,the research on the key technologies such as spectrum sensing,spectrum allocation,spectrum sharing,and spectrum mobility management in the cognitive vehicular net-works has achieved fruitful results.Considering the large time delay and energy consumption issues that widely existing in the practical application of these key technologies,this disserta-tion,which is supported by the National Natural Science Foundation of China project ”Research on Key Technologies of Internet of Vehicles”(project number: 61771126),explore the spec-trum prediction technology to effectively solve the time delay and energy consumption issues,and improve the overall performance of cognitive vehicular networks.The main contributions of this dissertation are as follows.1.Aiming at the problem of low efficiency of spectrum sensing in the interweave spectrum sharing mode,a cooperative spectrum prediction method is proposed to help deal with the sensing channel selection issue,and a new cooperative spectrum sensing scheme is proposed to alleviate the cooperation delay as much as possible.First,the neural network-based local predictors are trained separately based on the historical spectrum sensing results at multiple locations,and the genetic algorithm is used to optimize the initial parameters of each local predictor,so as to ensure the prediction accuracy of each local predictor.Then,using a weighted selection combining scheme to process the prediction results of each local predictor,and make full use of the spatial diversity among high-performance local predictors for cooperative spectrum prediction,thereby effectively improving the prediction performance of the licensed channel states in the interweave spectrum sharing process.Additionally,a new cooperative spectrum sensing scheme is designed on the premise of meeting the requirement of sensing accuracy and reducing the cooperative vehicles selection frequency and the num-ber of cooperating vehicles as much as possible,so as to improve the efficiency of spectrum sensing.Finally,simulation results show that,the proposed cooperative spectrum prediction method can provide higher channel state prediction accuracy,help with channel selection for spectrum sensing,reduce the time delay as well as energy consumption,and effectively improve the throughput of the cognitive vehicular networks.2.Aiming at the problem that the existing spectrum prediction methods cannot ac-curately capture the complex inherent dependency and heterogeneity among the spec-trum data to realize high-quality spectrum prediction in the underlay spectrum shar-ing mode,an end-to-end deep-learning-based model is proposed to make full use of the multi-dimensional dependencies and some particular time-domain characteristics of the collected spectrum data in the cognitive vehicular network to effectively solve the prob-lem.First,through the analysis of the real-world spectrum data,it is found that strong corre-lation characteristics exist in the time dimension,frequency dimension,and space dimension.Closeness,daily period and weekly trend characteristics are also found in the time dimension.Then,based on the observed characteristics,a deep learning based multi-dimensional spectrum prediction model is designed.Specifically,the most advanced convolutional long short term memory model Pred RNN is adopted to effectively model the spatial,temporal and spectral de-pendencies of the spectrum data.Three components based on the Pred RNN are designed to model the temporal closeness,daily period and weekly trend,respectively.And their outputs are dynamically aggregated to generate a final output.Finally,a large number of experiments are carried out based on real-world spectrum data to verify the effectiveness of the proposed multi-dimensional spectrum prediction method.It is shown that the proposed multi-dimensional spectrum prediction model outperforms the other spectrum prediction methods in terms of both prediction accuracy and computation and time complexity.It is also shown that the proposed multi-dimensional spectrum prediction model is promising to be applied to spectrum prediction with different prediction ranges.3.Aiming at the low quality problem of the collected historical spectrum sensing data in the cognitive vehicular networks,a new evaluation index termed maximum pre-dictability is introduced to analyze the impact of missing data and anomalies on the per-formance of spectrum prediction.A new data recovery method based on tensor comple-tion is proposed to effectively improve data quality and alleviate the impact of missing data and anomalies on spectrum prediction.First,by introducing the concept of maximum predictability,the impact of historical spectrum sensing data quality on spectrum prediction per-formance is analyzed from the information theory perspective.It is found that missing data and anomalies will not only reduce the predictability of licensed channels,but also affect the pre-diction accuracy of each spectrum prediction method.Then,based on the observations,a new data recovery method is proposed to effectively solve this problem.According to the spectral correlation,temporal correlation,and time regularity of the historical spectrum sensing data,a hankelized spectrum tensor is formulated considering missing data and anomalies.CAN-DECOMP/PARAFAC(CP)decomposition is introduced to decouple different interdependent tensor modes,and a robust online spectrum data recovery algorithm based on the alternating direction method is proposed to recover missing data and separate anomalies.Finally,a large number of experiments are carried out using real-world spectrum observation data to verify the proposed spectrum data recovery method.Experimental results show that the method is su-perior to existing data recovery algorithms in terms of recovery accuracy and computational efficiency.Besides,it is also shown that the proposed spectrum data recovery method not only improves the maximum predictability of licensed channels,but also improves the prediction accuracy of various spectrum prediction algorithms.4.Aiming at the low update speed and poor data quality problem of the commonly used spectrum maps in cognitive vehicular networks,a robust online spectrum map pre-diction method is proposed by effectively integrating tensor completion and time series forecasting techniques to improve the efficiency and effectiveness of the spectrum maps.First,considering that in the real-world cognitive vehicular networks,it is unrealistic to update the spectrum map frequently,and the slowly updated spectrum map is difficult to meet the needs of the vehicles to dynamically access the spectrum resources of licensed channels.To this end,a solution of spectrum map prediction is proposed.It is proposed to construct a three-dimensional tensor by integrating an efficient time series technique to represent the spectrum map,and the spectrum map prediction problem is transformed into a joint optimization problem of tensor completion and subspace learning.By fully considering the widely existing missing data and anomalies in historical spectrum sensing data,as well as the high computational complexity and time complexity of the traditional tensor completion methods,a new online tensor completion algorithm is designed based on the multiplier alternating direction method to efficiently recover the formulated tensor and achieve a complete and accurate spectrum map in advance.Finally,a large number of experiments are carried out based on real-world spectrum sensing data to verify the effectiveness of the proposed spectrum map prediction method.Experimental results show that this method is superior to the existing spectrum map prediction methods in terms of both prediction accuracy and computational efficiency,and can effectively alleviate the impact of missing data and anomalies on the effectiveness of spectrum map prediction.