Research On Road Traffic State Reconstruction Method In Perspective Of Knowledge-Aware

The 14 th Five-Year Plan for the Development of Modern Comprehensive Transport System emphasizes the need to accelerate the in-depth promotion and application of intelligent technologies and strengthen the refined control of urban transportation by promoting the deep integration of new technologies(big data,artificial intelligence,etc.)with the transportation industry.Road traffic state reconstruction can provide data assurance for intelligent travel services and is the core key to the accurate and active control of intelligent transportation systems.Traffic data is a prerequisite of road traffic state reconstruction.Compared to traditional fixed detector data,mobile data(such as floating car GPS data)contains a wide range of continuous traffic flow information,and has gradually become an important data source in the field of intelligent transportation in recent years.Mobile sensors can extensively sample information in large-scale,complex,and dynamic urban environments,providing a new possibility for accurate reconstruction of road traffic state.However,two problems seriously restrict the reconstruction of road traffic state:(1)currently,the penetration rate of mobile sensors on the road is still at a low level,and the sampling interval is set relatively large.The traffic state represented by mobile data is sparse.At this point,little disturbances in the data can cause large deviations in the traffic state estimation results.Therefore,suppressing error interference in sparse traffic data is one of the urgent issues to be addressed;(2)different from aggregated sensors,the traffic state information sampled by mobile sensors has deviations.According to the Shannon sampling theorem,only interpolating or filling traffic data cannot obtain an accurate traffic state under sparse and biased sampling conditions.Traffic state reconstruction from sparse data becomes an ill-posed problem.Therefore,designing a traffic state generation and identification mechanism to improve the robustness of deep learning methods is also an urgent issue to be addressed.Although mobile data is sparse,the carrier of traffic data,traffic flow,contains various types of traffic knowledge in different traffic scenarios.Traffic knowledge is a beneficial complement to sparse data,which can reduce the solution space of traffic state and guide algorithms to generate continuous and accurate traffic state.To obtain accurate traffic state from sparse data,the paper proposes a novel knowledge-aware approach for traffic state reconstruction based on typical GPS mobile data.The knowledge-aware approach digs traffic knowledge from observed data and uses knowledge to guide the generation of traffic state.This approach can both improve the accuracy and interpretability of deep learning.The specific work of this paper is as follows:First,a unified traffic knowledge representation and causality description is constructed.The existing structures of traffic knowledge are chaotic,and the knowledge semantics are overlapping and fuzzy.The algorithm of traffic state reconstruction is easy to misuse traffic knowledge.Therefore,it is necessary for knowledge-awaring traffic state reconstruction to construct a traffic scenario knowledge base that orderly classifies and uniformly represents traffic knowledge.Considering the demand for traffic state reconstruction,the paper divides traffic knowledge into three parts: knowledge of traffic scenario elements,traffic flow dynamic law,and traffic knowledge coupling rules.The paper establishes a comprehensive and unambiguous representation of three types of knowledge.Meanwhile,a causal graph is introduced to characterize the causal relationship between traffic knowledge and state.The causal graph takes traffic scenario elements as nodes and traffic flow dynamics as edges between nodes,providing a computational carrier for traffic state reconstruction to infer traffic knowledge.Second,a traffic scenario knowledge-aware model based on causal relationship is established.Traffic scenarios may contain a variety of knowledge to describe different spatial-temporal evolution laws of traffic flow.This knowledge is time-varying and uncertain.To adaptively infer the prior knowledge contained in the current traffic scenario,graph embedding technology and time series mining technology are used to mine the causal relationship between traffic scenario elements and traffic state.These technologies extract the causal effect of each piece of knowledge on the dynamic evolution of traffic flow,and estimate the probability distribution of prior knowledge in the traffic scenario.Then,the model uses the estimated knowledge to infer the traffic state and evaluates the distance between the discovered knowledge and the traffic state.Finally,a traffic knowledge causal discovery network based on an encoder-decoder structure is constructed to realize traffic scenario knowledge-aware.The effectiveness of the knowledge-aware model is tested on the examples of the queuing-release law of traffic flow at urban intersections and the propagation law of road traffic flow.The experimental results show that the knowledge-aware model can quantify the correlation between traffic scenario elements and traffic state.The average absolute error of traffic state estimated by the discovered knowledge is 2.19 km/h,which proves that the model can provide accurate knowledge for road traffic state reconstruction.Third,a knowledge-guided error suppression method for traffic data is proposed.To improve the reliability of existing self-supervised learning algorithms in suppressing sparse data errors,prior knowledge that adjacent data points conform to specific probability distribution is introduced.Under the maximum posterior estimation method,a knowledge-guided optimization problem of error suppression and a neural network solution algorithm is proposed.The error suppressor of the neural network analyzes the time characteristics of the traffic data through the Long Short-Term Memory network(LSTM)network and performs preliminary error suppression on the sparse data.The knowledge checker substitutes the output result of the suppressor into prior knowledge for verification.The test function in the Sobolev space is introduced to relax knowledge posterior constraints on error suppression when prior knowledge exceeds the applicable range,and a neural network is used for knowledge calibration and solution.Finally,a knowledge-guided error suppression method for traffic data is implemented.The experimental results show that the proposed method can remove 25.82 db error interference from sparse traffic data,and the structural similarity between the removed error and real traffic features is only 0.03.The results prove that the proposed method can provide accurate traffic data input for road traffic state reconstruction.Fourth,a traffic state reconstruction method based on traffic scenario knowledge is constructed.To reconstruct accurate traffic state,the paper designs a knowledge discovery-identification mechanism,which embeds the traffic knowledge-aware model into the traffic state calculation process and constructs a traffic state reconstruction method based on traffic scenario knowledge.The method designs a traffic adaptive convolutional neural network to generate complete traffic state from sparse data and identifies the posterior knowledge in the generated state.According to the principle of knowledge discovery-identification mechanism,the similarity between the posterior knowledge and the prior knowledge output from the knowledge-aware model is used to adjust the convolutional network.The convolutional network will re-generate a traffic state more consistent with the knowledge semantics.The paper uses typical GPS mobile data from taxis and buses in Changchun City to verify the performance of traffic state reconstruction.The proposed method is compared with advanced algorithms such as LSTM and Generative Adversarial Network(GAN).Experimental results show that the proposed algorithm performs well under sparse data and error interference with a certain reconstruction accuracy.The reconstructed traffic state can present continuity and fluctuation characteristics of traffic flow at the same time.Aiming at traffic state reconstruction under sparse data,the paper analyzes the causal relationship between traffic knowledge and traffic state and designs intelligent algorithms for traffic scenario knowledge-aware,error suppression in sparse traffic data,and road traffic state reconstruction.These efforts ultimately achieve continuous and refined calculation of road traffic state under knowledge-aware.The research is the progression of existing methods in traffic state mechanisms mining,which can improve the accuracy and interpretability of traffic state estimation under sparse data.Besides,the research result can provide data and service guarantees for traffic control and intelligent travel services.