Research On Outlier Detection And Prediction Technology For Spatio-temporal Trajectory Data

Based on coordinate position information,spatio-temporal trajectory data can objectively reflect the motion law of moving objects.It is of great application value to explore potential motion patterns in trajectory data,such as traffic route optimization,military target positioning,weather cloud warning and prediction,etc.At present,the research on spatio-temporal trajectory data is mainly focused on the traffic field with road network constraints,and the research on the movement trajectory of free space without road network constraints is less.Trajectories such as hurricanes or typhoons are typically unrestricted.The trajextory without road network restriction has the characteristics of small amount of annotation,strong mutation,high randomness,heterogeneous data source,cold start and sparse data,which brings difficulties and challenges to the study of spatio-temporal track data.This paper focuses on outlier detection and prediction technology of spatio-temporal trajectory without road network restriction,and proposes several anomaly detection and prediction methods by using machine learning and deep learning methods.The main research contents are as follows:(1)In order to quickly and effectively mine abnormal trajectories from a large number of trajectory data without abnormal tags,a two-level outlier trajectory detection algorithm based on segmenting clustering is proposed in this paper.Firstly,the state discrete index is proposed in segment stage to compress the trajectory by considering the changes in time and space.Secondly,the trajectories are clustered into normal pattern and abnormal pattern by combining time distance and space distance.Finally,a two-stage trajectory outlier detection algorithm is used for outlier detection.The first level of coarse-grained outlier detection is used to find the abnormal trajectory segment,and the second level of fine-grained outlier detection is used to find the minimum particle anomal subtrajectory.(2)Traditional clustering methods produce a large number of parameters and cannot meet the real-time requirements.In this paper,variational autoencoder is applied to trajectory outlier detection,and a trajectory outlier detection method based on variational autoencoder is proposed.Firstly,the trajectory sequences with different lengths are divided into subtrajectory sequences with equal length by sliding window as the input of variational autoencoder.Secondly,long short-term memory(LSTM)is used as the basic unit of variational autoencoder neural network for unsupervised learning to train the trajectory reconstruction model.Finally,the distance between the trajectory and the original trajectory is reconstructed from parallel distance,vertical distance and angle distance.If the calculated distance is greater than the given threshold,it is defined as an outlier.(3)When there is multi-source information such as structured sequence data and images in the trajectory data,the spatio-temporal characteristics of the trajectory can be fully mined by using the diversity of the information to improve the accuracy of trajectory prediction.Therefore,a trajectory prediction model based on deep multimodal fusion and multitask generation is proposed.The model mainly includes deep feature fusion module and multitask generation module.Firstly,a multimodal trajectory is divided into multiple multimodal subtrajectories.The LSTM Neural network and 3D Convolutional Neural Networks(3D CNN)extract the spatio-temporal features of the trajectory sequence and the dynamic features of the trajectory image respectively and achieve feature fusion.Thirdly,the fusion features of each multimodal subtrajectory are deeply fused to form a deep feature fusion module.Finally,multitask learning is applied to trajectory coordinate prediction to predict longitude and latitude at the same time to form a multitask generation module.(4)In real scenes,most trajectory data are based on latitude and longitude sequence data.It is more difficult to predict trajectory in single source condition than in multi-source condition.Therefore,taking advantage of deep learning in complex nonlinear modeling,this paper proposes a trajectory prediction method based on long short-term memory network and Kalman filter.Firstly,the trajectory sequence is grided and one-hot coded.At the same time,the trajectory vector generated after coding numeric data and classification data is input into LSTM to predict the region where the trajectory appears.Secondly,based on the simple LSTM trajectory prediction algorithm,the candidate set of prediction results is established,and an improved LSTM trajectory prediction algorithm is proposed by conditionally screening the candidate set.Finally,combined with Kalman filter,the trajectory predicted by the improved LSTM trajectory prediction algorithm is filtered to improve the accuracy of prediction results.(5)Deep learning methods need to be based on rich features.In order to extract richer temporal and spatial features from simple trajectory sequences,a trajectory prediction model based on deep feature representation is proposed.The model includes trajectory data preprocessing layer,deep feature representation layer and trajectory prediction layer.Firstly,a long trajectory is divided into multiple subtrajectories by a sliding window in the data preprocessing layer.Secondly,bi-directional long-short Term Memory(Bi LSTM)is used to extract the temporal and spatial features of subtrajectory sequences,and 3D CNN is used to extract 3D dynamic features after the subtrajectory sequences are transformed into images,and then feature fusion is performed at deep feature representation layer.Finally,at the trajectory prediction layer,the self-attention model is used to assign different weights to the features extracted from different subtrajectories,and Bi LSTM neural network is used to predict longitude and latitude simultaneously.The three-layer model is used to fully excavate the temporal and spatial characteristics of trajectory and carry out the trajectory application research and analysis.(6)In this paper,hurricane and typhoon are taken as the specific application scenarios,and the trajectory outlier detection and prediction techniques are integrated to achieve real-time trajectory outier detection and prediction.Firstly,the trajectory outlier detection method based on variational autoencoder is used to determine whether the trajectory is abnormal.Then,the trajectory prediction model based on deep feature representation is used to predict the movement trend of abnormal trajectory.Finally,the outputting trajectory anomaly detection and prediction results provide effective decision support for meteorological disaster warning.