Research On Vessel Trajectory Analysis Method Based On Machine Learning

With the number of vessels increasing and maritime traffic getting busier,it is urgent to improve regulatory efficiency and ensure the safety of vessel navigation.The data from the Automatic Identification System(AIS)provides important data support for maritime supervision and analysis of maritime traffic status.Vessel trajectory analysis is a collection of analytical methods based on a large amount of vessel AIS data,aimed at fitting(restoring)abnormal and missing vessel trajectories,establishing predictive relationships between time and trajectory,and extracting vessel trajectories within a specific area based on a collection of trajectories.Vessel trajectory analysis based on AIS data can provide data support for improving regulatory efficiency,maritime safety warning,maritime timeliness analysis,port and its affiliated facilities construction planning,channel design,etc.This thesis mainly studies the trajectory analysis methods,including missing trajectory fitting,trajectory prediction,and trajectory extraction,aiming to provide more accurate and comprehensive data support for maritime supervision and maritime traffic state analysis.At present,there are still some problems in the above three aspects of research that need to be further improved.For example,the missing trajectories have a low fitting degree,the distribution of predicted trajectory points is random,and the prediction error fluctuates greatly.The trajectory of hub transportation nodes cannot be extracted,and many noise exist.To solves these problems,this thesis proposes a series of vessel trajectory analysis algorithms based on machine learning,as shown in the following.1.This thesis proposes an improved missing trajectory fitting algorithm based on bidirectional Gated Recurrent Unit(GRU)network,to address the problem of low restoration performance and large error of traditional missing trajectory fitting algorithm.The algorithm defines the trajectory acceleration through the trajectory at both ends to be fitted,constructs a bidirectional linear expression based on the acceleration,and combines the expression with the GRU network.The algorithm in this thesis uses a two-way simultaneous advancement method to train known trajectories to gradually reduce the length of the trajectory to be fitted,thereby reducing the fitting error value.This thesis conducts multi-algorithm comparative experiments based on multiple perspectives such as algorithm performance,trajectory type,and navigation scenarios.The experimental results show that the algorithm in this thesis is superior to other algorithms,and the minimum fitting root mean square error of a single cargo vessel trajectory is 0.0117.2.This thesis proposes a vessel trajectory prediction algorithm based on GRU network and attention mechanism,to address the problems of random distribution of predicted trajectory points and insufficient feature representation of key inflection points(the point with a large change in the trajectory movement trend before and after)in traditional vessel trajectory prediction algorithms.The algorithm constructs a Trajectory Direction Vector(TDV)from the features of the vessel’s course and speed,to determine the direction of the trajectory change trend,thereby constraining the direction of trajectory convergence and reducing the fluctuation of prediction error.The algorithm combines TDV with the weight in the attention mechanism,and trains the weight of the trajectory point with a large change of trajectory trend through weight,thereby improving the correlation of adjacent trajectories,and fuses the weight with the two-way GRU network and constructs a fully connected layer that associates time and trajectory points,so as to realize sliding trajectory prediction.This thesis conducts multi-algorithm comparison experiments based on algorithm performance,multiple time periods,and navigation scenarios.The experimental results show that the algorithm in this thesis is superior to other algorithms,and the error is only 439.68 meters and the fluctuation range is less than 50 meters in the 120-minute prediction task.3.This thesis proposes an algorithm for sailing route extraction based on fast density clustering and outlier factor estimation,to address the problem of low precision and many noise in traditional sailing route extraction algorithm.The algorithm improves the similarity judgment condition in the density clustering algorithm,so that it can distinguish the trajectory direction.Based on the idea of hierarchical distance division,the algorithm improves the traversal method of trajectory points and clusters,making it change from the original point-by-point traversal to merging between clusters,thereby reducing the computational time complexity.The algorithm calculates the outlier factor and estimates the trajectory points with high outlier degree and low probability distribution through the kernel,and eliminates noise by mapping probability and trajectory plane.This thesis conducts comparative experiments including clustering performance,algorithm running time and channel noise reduction.Experimental results show that the algorithm has good accuracy and robustness.Under the same experimental data,the noise data in the clustering results of this algorithm is 14.34% lower than other algorithms,and its completeness and homogeneity are 15% higher than other algorithms.