Research On Methods Of Crowdsensing Data Processing Based On Tensor Analysis

With the development of social economy,the problem of traffic congestion becomes more and more serious.Traffic flow prediction can provide important reference for travelers and relieve traffic pressure to a certain extent.The prediction systems require sufficient data support,and emergence of crowdsensing technology provides a new way to acquire traffic data.At the same time,the analysis of the obtained data is also an important part of the crowdsensing system.In most cases,the data actually obtained is missing and abnormal for some reason.It has been a research hotspot in this field to make better analysis and prediction in the case of incomplete data.Starting from the analysis and modeling of traffic flow data,this paper proposes a traffic flow prediction method based on tensor analysis.The main contributions are as follows:(1)First analyze the multimodal correlation of traffic data,and a static tensor model covering the four dimensions of week,day,time and space is constructed.Using the low rank approximation of the Tucker-ALS algorithm,it is proved that the tensor model is low rank and satisfies the low rank hypothesis of the tensor filling algorithm.Based on the static tensor model,a series of tensors with the same size of each dimension are constructed by reducing the time dimension and setting the sliding time window.It meets the background of real-time update of traffic flow data.A dynamic tensor flow model is formed by arranging tensors of these dimensions in chronological order.(2)Based on the dynamic tensor model,an adaptive n-rank dynamic tensor analysis algorithm is proposed.It consists of two parts: updating the projection matrix and predicting the unknown data.In the updating part of projection matrix,an adaptive n-rank algorithm is introduced to calculate the optimal rank of input tensor in real time.The forgetting factor is set and the projection matrix is updated with the historical tensor and the real-time tensor.In the prediction part of unknown data,iterative alternating least square method is introduced to further optimize the projection matrix so as to obtain the optimal projection matrix.The real-time tensor is projected along the projection matrix to obtain the core tensor.Finally,the projection matrix and the core tensor are used to calculate the unknown time data in the real-time tensor.Experimental results show that the adaptive n-rank dynamic tensor analysis algorithm has better prediction performance than dynamic tensor analysis algorithm and other static tensor filling algorithms.(3)For the real-time tensor in the dynamic tensor model,since the time dimension contains the unknown time,there is a lack of column fiber no matter whether the obtained historical data is complete or not.This paper proposes a preprocessing method for missing data,which can recover the original structure of the tensor as much as possible.By introducing the tracking projection matrix algorithm,the input tensor is expanded into a matrix,and the columns of the matrix are rearranged according to the value of the norm.The first few columns with the largest norm are selected to update the projection matrix by using the tracking projection matrix algorithm,so as to avoid the diagonalization process.Based on this,a flow tensor analysis algorithm is proposed.Experimental results show that the flow tensor analysis algorithm reduces the complexity of the algorithm while maintaining almost the same predictive performance as the adaptive n-rank dynamic tensor analysis algorithm.