Research On Some Key Technologies Of Big Data In Vehicular Cyber-Physical Systems

With the advent of big data and intelligent era,intelligent assisted driving is changing the way of people's travel.Alleviating traffic congestion and improving travel safety is significant for smart travel.The vehicular cyber-physical system(CPS)is a deep fusion system that integrates information,computing,network and physical environment to realize real-time awareness,dynamic control and information service of the system.Traffic flow prediction,as a means of providing effective information services for path planning in automobile driving,has become an important issue in vehicular CPS cognitive computing.This thesis takes cognitive computing in the vehicle CPS as an objective and takes traffic flow big data as a research example,focusing on tensor-based data representation,the cognitive method,and the allocation strategy of computing resources for parallel applications in the vehicular CPS.This research provides real-time and accurate information services for personal intelligent assisted driving in vehicular CPS from the perspectives of data representation,cognitive model and parallel computing.Experiments are performed and the results validate the proposed methods.The major work and contributions of this thesis are summarized as follow.(1)For the cognitive computing problem of big data in the vehicular cyber-physical system,cognitive computing in the vehicular CPS is discussed and then the methods of traffic flow prediction are summarized by taking traffic flow prediction as a research example.Aiming at the problems brought by CPS big data,this thesis combines them with deep learning methods and reviews the research progress of deep learning cognitive computing in the big data environment in recent years.Firstly,data representation,cognitive model and parallel computing of deep learning are focused.Secondly,the application and challenge of deep learning cognitive computing for the vehicular CPS are summarized.Based on this,this thesis studies the key points such as data representation,cognitive model and parallel computing of cognitive applications for vehicular CPS big data.(2)For the diversity problem of vehicular CPS big data,the tensor-based data representation is constructed and the tensor-training deep computation model is proposed.Taking the application of traffic flow prediction as a research example,a tensor data representation and a deep convolutional network prediction model are firstly constructed,and then a tensor-based deep computation model TDCN(Tensor-based Deep Convolutional Network)is proposed.However,computing resources in the vehicular CPS are limited and the calculation of the tensor-based deep computation model is complex.In order to reduce the computational complexity of the model,a deep computation model with Tucker decomposition for vehicular CPS prediction is proposed by introducing Tucker decomposition into the deep computation model.The experimental results on the TaxiBJ data set verify the effectiveness of the proposed models in terms of accuracy,parameter reduction and speedup.(3)For the uncertainty problem of big data in the vehicular CPS,taking traffic flow big data as a research example,the methods of traffic flow prediction based on fuzzy deep convolution network,namely,FCNN(Fuzzy Convolutional Neural Networks)and FDCN(Fuzzy Deep Convolutional Networks),are proposed.The proposed methods introduce fuzzy theory into deep learning model to reduce the influence of data uncertainty.The fuzzy rules can be adaptively generated by constructing a fuzzy deep convolutional network model and designing a self-learning algorithm.The FDCN model based on the deep residual network explores the spatial-temporal correlation of traffic flow to improve the accuracy of prediction and avoids the over-fitting problem in the process of model training.The experiments are performed from the aspect of model convergence,model structure optimization and prediction accuracy.Experimental results show that the proposed method has superior performance compared with state-of-the-art approaches.(4)For real-time performance requirements of the vehicular CPS big data application,computing resource allocation strategies of parallel applications in the heterogeneous cloud system for minimizing execution time and minimizing execution cost are researched.The algorithm of minimizing the execution time of budget-constrained parallel applications MSLBL(Minimizing the Schedule Length using the Budget Level)and the algorithm of minimizing the execution cost of deadline-constrained parallel applications DUCO(Downward Upward Cost Optimization)are proposed.The proposed method first construct the system model,then decompose the research problem into two sub-problems,and finally transform the problem and propose heuristic algorithms.The MSLBL defines a parameter of budget level to fairly preassign cost to each task,and the DUCO uses heuristic algorithms that optimize the execution cost from downward and upward perspectives to allocate computing resources to tasks.The experimental results verify the effectiveness of the proposed methods.In summary,this thesis focuses on the real-time and accurate cognitive computing problem of big data in the vehicular cyber-physical system,analyses the vehicular CPS big data representation,cognitive algorithms and parallel application methods,and proposes corresponding models,algorithms and strategies.The research results bear both theoretical and practical significance for improving the cognitive accuracy and real-time performance of the vehicular CPS.