| With the rapid development of China’s international trade,particularly since the proposal of the Maritime Silk Road construction in 2013,the volume of maritime shipping has increased sharply,putting tremendous pressure on maritime traffic monitoring and management.In order to ensure the safety and efficiency of maritime transportation,it is necessary to effectively monitor and manage the vessels’ sailing trajectories and timely detect any abnormal trajectories.In recent years,with the widespread adoption and mandatory implementation of Automatic Identification System(AIS),maritime authorities have access to massive amounts of AIS data,providing a richer data foundation for vessel trajectory anomaly detection.However,due to the complexity of AIS data and the variability of maritime environments,identifying abnormal trajectories from large amounts of data and designing suitable anomaly detection algorithms to meet the demands of different scenarios have become pressing challenges for maritime authorities to tackle.To address this challenge,this paper proposes a deep learning-based approach for ship trajectory anomaly detection.The specific work is as follows:The third chapter of this paper preprocesses the original AIS data and constructs the AIS dataset.This part of the work includes: handling abnormal data;selecting appropriate interpolation methods to interpolate missing values;resampling the original trajectory at the same latitude interval to obtain a new trajectory with the same number of trajectory points;generating corresponding ship trajectory route image sets;manually labeling the trajectory data according to the "General Rules for the Ship’s Route System";dividing the dataset into training,testing,and validation sets according to a certain proportion,and completing the construction of the dataset.The third chapter followed by constructing a ship abnormal trajectory detection model based on multi-modal feature fusion.This part of the work includes: designing a ship trajectory visual feature extractor based on Res Net18 network,AIS data text feature extractor based on Bi-LSTM,and LSTM-based Feature fusion network.The final model will output the predicted sequence,and complete anomaly detection by measuring the error between the real sequence and the predicted sequence.The fourth chapter designs an experiment to verify the feasibility of the method proposed in this paper.The specific work is as follows: build the environment required for the experiment;determine the experimental evaluation criteria;determine the hyperparameters used by the model through data obtained from a large number of experiments;design a comparative experiment to verify the method proposed in this paper Effectiveness;A large number of comparative experiments have been carried out with traditional methods.The method designed in this paper is compared with some traditional methods in the same data scenario.The experimental data shows that when the four indicators of the confusion matrix are used as the score,the score obtained by the method in this paper is significantly better than that of the traditional method.It proves that the method proposed in this paper has certain advantages compared with traditional methods.The fifth chapter designs and implements a prototype system for ship track anomaly detection based on sufficient demand analysis and ship track anomaly detection model based on feature fusion.The prototype system supports functions such as user registration and login,uploading AIS trajectory data,performing anomaly detection and viewing detection results,etc.,and can assist maritime staff in real-time judgment on the abnormal state of ship trajectory.In a word,this paper proposes a ship track anomaly detection method based on feature fusion and realizes it.This method can effectively monitor ship track and detect abnormal situations in time,which provides strong support for maritime safety supervision and has practical application value and promotion significance. |
