Research On Recovery Of Missing Data Of Urban Traffic Trajectory Based On Deep Learning

Missing data is a problem that exists in many fields in the context of the big data era.Many experts and scholars have been actively researching the processing of missing data in the fields of transportation,electricity,and medicine,etc.The research also found that the missing data recovery method in the field of transportation can be extended and applied to other fields.Traffic trajectory data is a kind of traffic big data,providing an indispensable data base for many potential fields including intelligent transportation systems,urban traffic planning and driverless technology.High-quality trajectory data must ensure its integrity and validity.However,during data collection and transmission,it is susceptible to communication instability,positioning errors and low sampling rates,which make data missing and erroneous.Such incomplete data can lead to information damage and bring adverse effects to various subsequent work and research related to trajectories.It can be seen that the problem of missing data is still a difficult problem to be solved in the current traffic field.Therefore,the realization of traffic trajectory data recovery in the complex urban road environment is of great practical significance to the development and construction of cities.In this thesis,two methods for recovering missing trajectory data are constructed based on deep learning using the trajectory dataset provided by the Department of Transport of Fujian Province as the research object,and the effectiveness of the methods is demonstrated through a large number of experiments.The main work of this thesis includes:(1)Urban Road Network Modeling and Dataset Pre-processing.The urban area of Fuzhou city is selected as the study area,and the urban road topology is modeled based on the real road network and using graph theory.The real trajectory dataset is used as the research object,and the distribution characteristics of trajectory points are analyzed in depth to provide a reference for the establishment of feature learning methods.Finally,in order to improve the data quality and adapt to the trajectory recovery task,the data are pre-processed to provide a good data base for the subsequent study.(2)A missing trajectory recovery method based on a sequence encoder-decoder model is proposed.In the model,the Seq2 Seq framework is constructed by LSTM neural network,the construction of trajectory vectors is realized by trajectory embedding,and the spatiotemporal attention mechanism is introduced to enhance the learning ability of the model on the spatiotemporal features of trajectories.The feature construction of exogenous influencing factors is realized through a potential factor module to enhance the reliability of trajectory recovery in specific scenarios.Numerical experiments based on real trajectory datasets and real road network information in Fuzhou,and the results demonstrate the effectiveness of the method in various data-missing scenarios.(3)In order to solve the problem that the original Seq2 Seq framework relies on the already observed sequences and cannot adjust the never observed road segments to affect the recovery of the missing locations,the missing trajectory recovery method based on generative adversarial network is further proposed.Based on the seq2 seq framework,a generative adversarial network is further constructed to realize the adaptive learning of the overall distribution characteristics of trajectory sequence.At the same time,a reinforcement learning mechanism is introduced to train the generative network in the model by the policy gradient method,so as to avoid the model falling into the local optimum.And through an evaluation network to evaluate the work of the discriminative network to improve the efficiency of model training.The experimental results show that the recovery performance of the model has been further improved.