Research On Prediction Method Of Ship Location Based On Deep Recurrent Neural Network

The 21 st century is known as the "century of the sea".The transportation,trade,resources,and strategic deployment of numerous countries are all impacted by concerns of navigation safety and sovereignty over territorial seas.As a result,whether in ocean journey scenarios or offshore activities,it is necessary to implement effective monitoring and management of all ships.The goal of location prediction is to reliably and efficiently anticipate the upcoming trajectory of the target ship,which is the foundation for abnormal state detection,route control,risky behavior discrimination,and many other ship-related tasks.This research will be critical in collecting real-time ship dynamic information and ensuring regional security.From the prospective of ship location data features,this thesis explores applications of RNNs in spatiotemporal sequence prediction tasks,which both considers conventional ocean situations and offshore scenarios with high uncertainty.Different prediction methods will also be proposed for ultralong dense sequences and low-rank sparse sequences.The main research work includes:Through the visual analysis and the preprocessing of AIS ship dataset,restoring the real ship activity trajectory and summarizing the distribution law of the data.Then,integrating ship state attributes to complete data cleaning,sequence coding,and missing value supplement works in the context of mission needs and characteristics.In addition,according to the experimental objectives,AIS ultra-long dense sequences dataset and AIS low-rank sparse sequences dataset are constructed respectively for model verification and evaluation.This thesis designs Spatiotemporal Window Recurrent Neural Network for ultra-long dense sequences.The model uses shift windows to transform long-term dependency issues into the feature extraction problem of local sequences,from which the optimization method of capturing spatiotemporal and velocity information in the continuous time period of the window also further benefits the accuracy.In view of the disadvantage that the traditional RNN cannot be calculated in parallel,this model effectively improves the computational complexity based on a multi-head attention framework.The experimental results show that this model can overcome the gradient vanishing problem of long-term dependency sequences,and achieve efficient performance by combining the advantages of prediction accuracy and computational efficiency.This thesis proposes a Self-attention based Bidirectional Gated Recurrent Unit Network for lowrank sparse sequences.Considering the impact of data sparsity and omission on target behavior pattern mining,this model discovers semantic connections between locations in a bidirectional feature propagation structure,and uses dynamic data augmentation methods to share similar historical feature dependencies.Inspired by the human visual mechanism,this model adds a self-attention weighting layer that focuses on the source sequence,which can be used to screen and distinguish important location features.Several groups of experiments prove that this method can well alleviate the impact of data sparsity,and shows good stability and generalization on different types of datasets.