Research And Implementation Of A Multidimensional Data Simulation Tool For Maritime Targets Based On Generative Adversarial Networks

The efficient target identification technology based on marine sensor data such as automatic identification system(AIS),radar and sonar of ships is of great significance to guarantee marine security and maintain the development of maritime trade.Among them,multidimensional target data association and cumulative recognition is one of the research hotspots,and the algorithm performance of such research largely depends on the quantity and quality of the data collected by sensors.However,due to the limited public availability of data sets,the difficulty of collecting real data and the complexity of post-processing,how to obtain sufficient maritime target data for algorithm training and testing is a pressing problem in this field.There are already some data simulation works for maritime targets.However,most of the existing studies focus on data simulation in a single sensor dimension and do not consider both the scarcity of target sample data sets and the complexity of maritime target motion behavior.Therefore,this thesis proposes a multidimensional data simulation tool for maritime targets based on generative adversarial networks.Based on a small amount of real target sample data,the tool simulates a large amount of multi-dimensional target attributes and behaviors that are "similar to and different from" the real data through machine learning algorithms,which is used to support the group’s subsequent research on cumulative recognition and cognition algorithms for small targets at sea.To address the above issues,the main work of the thesis is divided into three parts.(1)The method of generating AIS track spatial data based on Conditional-WGAN-GP(abbreviated:C-WGAN-GP)is studied and implemented.In this thesis,the C-WGAN-GP network is implemented to solve the problem of generating a large amount of AIS track data.Using a small amount of real target AIS track data,it learns the spatial features of the data and follows its feature distribution to generate a large amount of basic AIS track spatial data.(2)A multi-dimensional data simulation tool for maritime targets is designed and implemented.Based on the above data generation method,this thesis combines sensor information content and equipment parameters,and uses data association inverse thinking achieve batch and real-time streaming generation of multidimensional simulation data of AIS,radar,sonar and optoelectronic.(3)For the abnormal behavior of the target,an abnormal event simulator is implemented and an abnormal event class framework is built to complete the data generation of the target abnormal behavior simulation and data scene customization,and provide the basis for the diversification of the subsequent abnormal behavior.Finally,the function and performance of the tool are tested.The results show that the data generation method is effective,and the multidimensional data simulation tool meets the requirements and has availability.The generated simulation data provides effective support for the training and verification of the target cumulative recognition algorithm and anomaly detection algorithm.