AIS Trajectory Similarity Computation Based On Graph Neural Networks

Ship trajectory similarity calculation is a method used to measure the degree of similarity between ship trajectories.Typically,in ship trajectory analysis,ship trajectory similarity calculation is an important foundational problem.A good method for measuring ship trajectory similarity is crucial for many subsequent tasks and applications,such as trajectory clustering,trajectory prediction,anomaly detection,and route planning.Therefore,it is necessary to conduct research on similarity calculation methods for ship trajectories.With the rapid development of the shipping economy and the popularity of onboard mobile devices,the collection and utilization of ship AIS data have entered a phase of rapid growth.How to utilize massive AIS data to serve intelligent shipping and shipping supervision has become a key problem that needs to be addressed.In the past,AIS trajectory similarity calculation methods based on traditional distance metrics and various improved algorithms have shown good performance.However,with the continuous increase in ship density in current waterway areas and the increasingly complex electromagnetic environment,existing classical AIS trajectory similarity calculation methods are difficult to meet the requirements in terms of performance,robustness,and computational efficiency.In recent years,some research efforts have attempted to incorporate advanced techniques based on deep representation learning into the calculation of similarity in traffic trajectories.However,there is a lack of explicit modeling methods for spatial dependency features in AIS trajectories using graph-based approaches or similar methods.This limitation has resulted in insufficient performance,robustness,and efficiency in AIS trajectory representation and similarity calculation.As a result,the development of AIS trajectory similarity calculation research has been constrained.In the aforementioned context,this study conducts research on the method and application of AIS trajectory similarity calculation based on graph neural networks.It constructs AIS trajectories as trajectory graphs to explicitly model the spatial dependency relationships in trajectories.By introducing the proposed trajectory graph computation frameworks and improved graph neural network models,accurate and robust trajectory representations are learned to address the aforementioned issues.The main innovations and research contents of this study include:To address the efficiency limitations of classical trajectory similarity calculation,a trajectory similarity approximation model learning framework called AISim is proposed.It leverages multiscale spatial features to improve the trajectory representation capability of the model.By combining metric learning methods,AISim achieves significant improvements in computational efficiency compared to classical methods.It also mitigates the challenge of integrating both trajectory and grid features in traditional trajectory gridding methods.The framework consists of three main innovations.Firstly,an improved water space pretraining method is introduced.It constructs a water structure graph using a historical trajectory database for pretrainin.Secondly,a multiscale trajectory graph construction method is proposed.It models different levels of features by simultaneously representing real trajectories and constructed grids.Thirdly,an improved heterogeneous graph neural network encoder is developed.It learns multiscale representations of the trajectory graphs through feature aggregation.Finally,traditional similarity metric methods are employed to calculate distances as supervision signals.The model is trained within the metric learning framework,establishing a universal learning model capable of learning various classical similarity measures.Experimental results demonstrate that the proposed method improves the computational efficiency of classical methods and outperforms advanced sequence learning models like Transformer by at least 5% in hit rate.It promotes research on approximate calculation methods for AIS trajectories.To address the performance degradation caused by the insufficient representation capability and robustness of existing AIS trajectory similarity calculation methods,a contrastive learning-based AIS trajectory similarity calculation framework called CLAIS is proposed.It calculates trajectory similarity by computing the Euclidean distance between extracted trajectory graph representation vectors.The framework introduces three main innovations.Firstly,the water space construction method mentioned earlier is improved by incorporating grid filtering for pretraining purposes.Furthermore,a parameterized combination trajectory augmentation scheme is proposed to enhance the model’s robustness in dealing with noise and data loss issues.Moreover,an improved graph neural network representation learning module is introduced to enhance the model’s perception of the spatial structure of trajectory graphs.Combined with contrastive learning methods,the model’s unsupervised representation learning capability is further improved.Additionally,an improved self-similarity experiment is proposed to conduct a more comprehensive and thorough evaluation of the CLAIS framework.The experiments demonstrate the robustness of the model when dealing with problematic trajectories and promote research on unsupervised AIS trajectory representation learning paradigms.To address the issue of underperforming similarity measurement in existing clustering models for anomaly detection,a similarity-based AIS trajectory anomaly detection framework is proposed.The framework builds upon the CLAIS framework with improvements.It learns representations of ship trajectories that incorporate features such as latitude,longitude,heading,and speed.By training the model,it acquires the capability to model various AIS trajectory features,thereby enhancing its ability to represent ship trajectories.In the proposed framework,unlabeled AIS trajectory data is input into the trained model to obtain trajectory representation vectors.The distances between these representation vectors are calculated,and in combination with the HBDSCAN clustering method,the optimal cluster selection is determined based on quantitative clustering metrics.This approach enables effective trajectory clustering,which is further visualized.Finally,a recurrent neural network trajectory predictor is trained based on the clustered normal models.By comparing the input trajectory with the predictor’s output,abnormal trajectory behavior can be detected.Experiments validate the proposed model’s ability to promote unsupervised trajectory clustering and its effectiveness in AIS trajectory anomaly detection.In summary,this dissertation proposes a trajectory similarity learning framework called AISim based on multi-scale spatial features,aiming to improve the computational efficiency of classical similarity calculation methods and promote trajectory graph representation learning.It also introduces a trajectory similarity calculation framework called CLAIS based on contrastive learning,which includes a parameterized combination trajectory augmentation scheme to enhance representation learning robustness,thereby improving the accuracy and robustness of unsupervised AIS trajectory representation learning.Furthermore,an improved CLAIS framework is proposed to simultaneously model heading and speed,and combined with HDBSCAN to construct normal trajectory patterns for detecting abnormal vessel behaviors,demonstrating the practical value of the proposed trajectory similarity calculation method in this dissertation.