Research On Data Collection And Data Imputation Based On CrowdSensing

With the continuous progress of society and the acceleration of urbanization,traditional urban management schemes can no longer deal with a series of problems caused by population growth effectively.In order to explore the new development models,cities must adopt new technologies.Nowadays,the development of new smart cities have been promoted by the popularization of intelligent terminal,vehicle GPS and other sensing devices as well as digital platforms.The technologies of smart cities driven by data take massive urban data as their core support,so the perception and acquisition technology of urban data is an important foundation for the realization of many functions.As a new generation of intelligent network,crowd-sensing uses group wisdom to accomplish large-scale urban perception tasks.How to ob tain the high quality city data while reducing the cost of collection is the one of research focus in this field.For urban traffic flow data,it is also an important part of urban functions realization to predict urban traffic conditions and population density to optimize urban services and public safety.However,the problem of missing data emerges in an endless stream.In order to obtain complete data accurately and extract more valuable information from them,it is urgent to solve the problem of missing data imputation with high quality and efficiency.For these reasons,the following research es are carried out on data collection and missing data imputation.(1)A distributed compressive data collection framework based on mobile crowd-sensing is proposed.When the participants move to some monitoring sites to gather data,the movement among these monitoring sites is si milar to the random walk among sensor nodes for data collection in wireless sensor networks.The participants' movements are essentially random and spatio-temporal,they take data from one location to another,and their trajectories actually provide valuab le random coverage for the sensing tasks.In a sense,the trajectories can completely cover the target area.This article proposes a distributed coding algorithm based on compressive sensing by using the movement of participants.Firstly,we simulate the movement of participants,through it,we can get the measurement vector of each participant,and then a binary measurement matrix can be obtained from multiple participant trajectories.The sensing data gathered by each participant in the target region duri ng a certain period of time are compressed into a measurement value.Finally,the original data is reconstruct by the iterative algorithm BAMP.In order to improve the quality of dat a reconstruction,a round-based data recovery algorithm is proposed.The experimental results show that the proposed scheme in Chapter Three can achieve better data quality under different decoding rates than the existing collection methods.(2)A missing data imputation model based on prediction model in crow d flow is proposed.Crowd flow is a typical traffic data,which is obviously with spatia l dependence and time-dependent.In addition,it is easy to be affected by the external environmental factors and shows a certain regularity of dynamic change.Deep learning with strong feature expression can comprehensively describe the attributes and factors mentioned above,so this paper designs a model to impute the missing crowd flow based on deep learning.In this model,the prediction network STRes Net and imputation network are combined together.First,the STRes Net is pre-trained.Then,according to mask matrix M,the missing data at the tim t is preliminarily imputated with the predicted value at the tim t,and the initial completed data are input into the imputation network for training.At this stage,the prediction network is frozen,and only the completion network's parameterss are updated.Finally,two networks are trained jointly to update the parameters of the whole model.The initial imputation with predicted values makes the distribution of missing data closer to the real data,which is beneficial to the learning and convergence of the network.The experimental results show that the proposed method in Chapter Four is superior to the existing method,and the combined training of prediction and completion network can further improve the accuracy of data prediction and imputation.