Structured Learning Of Ship Intelligent Collision Avoidance Decision-making Based On Seq-CGAN And Ship Handling Behavior Basic Splice

In the process of USV from use to overall popularization,it will go through a complex transition period of two or more types of ships encounter in the "unmanned to manned " mode,which is the bottleneck of maritime navigation safety research because the most influential factors are considered and the sea area environment is the most complex.In this paper,under the condition of "unmanned to manned" mode,the research on USV intelligent collision avoidance decision-making and key technologies are carried out.This article selects the topic of"Structured learning of ship intelligent collision avoidance decision-making based on Seq-CGAN and ship handling behavior basic splice" to study a new method for how to generate reasonable and effective ship intelligent collision avoidance decision-making.The effective combination of the Automatic Identification System(AIS)big data and the related theories of machine learning/artificial intelligence was expected to improve the practicality,effectiveness and rationality of ship intelligent collision avoidance decision-making.Based on AIS big data as the research foundation,this article around the ship intelligent collision avoidance decision-making and key technologies to analyze and study the related issues.In view of the AIS big data mining,ship encounter data extraction,ship trajectory key feature points recognition,ship navigation behavior prediction,ship handling behavior recognition and collision avoidance decision-making.As a whole,the research process adopts the way of "big data mining"-"big data analysis"-"big data learning" to study the ship intelligent collision avoidance decision-making,and strive to generate an anthropomorphic decision on collision avoidance that is close to conventional marine practices.At the same time,based on the AIS real ship trajectory data near Ningbo-Zhoushan port in 201 8,Tianjin Port in 2015,Laotieshan waterway in 2017 and Cheng Shantou waterway in 2017 to conduct of the verification and neural network training of the above methods.By using AIS big data,to overcome the disadvantages of low reliability of data sources,insufficient persuasion and difficulties in collecting basic data.Successful cases of collision avoidance at sea from AIS big data were extracted to establish ship collision avoidance behavior pattern library,and obtain reasonable and effective collision avoidance decisions by machine learning.The main research contents of this paper are:(1)A support vector classification(SVC)model was applied to construct the ship encounter azimuth map.The ship trajectory data of actual ship encounters was identified in the AIS data,the corresponding azimuth scatter points of ship encounters at sea were calculated,the big data mining results were combined with the "International Regulations for Preventing Collisions at Sea" to quantify accurately the values of each boundary of azimuth encounter map.The 8 encounter azimuth areas and 12 types of ship encounter patterns were inferred;(2)An improved variable-scale sliding window algorithm was proposed to construct a value measurement system for ship trajectory points,identify and extract key feature points with spatiotemporal characteristics from the AIS trajectory data.In this way,the overall AIS data quality was improved,the amount of calculation in subsequent research was reduced,and the hidden ship handling behavior in the trajectory data was mined;(3)The concept of "ship handling behavior basic(SHBB)" was proposed.Then the ship trajectory at the position of the ship key feature points was sliced,and the combination algorithm of T-distributed stochastic neighbor embedding(T-SNE)and spectral clustering was proposed to cluster the ship sub-trajectory segments by truncation.The 11 types of ship handling patterns were inferred;(4)A bidirectional long short-term memory recurrent neural network was constructed to predict the navigation behavior of ships,to predict the handling intention of other ships while sailing,which can improve the accuracy,effectiveness and priority of the ship intelligent collision avoidance system.(5)An Encoder-Decoder response network was constructed to perform seq2seq structured machine learning based on the ship collision avoidance behavior pattern library.By this way,ship encounter data which had been screened and identified was quickly learned,and ship collision avoidance decision-making by way of response was generated;(6)A sequence-conditional generative adversarial network(Seq-CGAN)was constructed by transforming the trained Encoder-Decoder network into conditional generative adversarial networks and combining with the proposed splicing rules of ship handling behavior basic to make the generated decision-making is more close to the common practice of marine navigation at sea and meet the requirements of good seamanship,improve the integration with manned ships,and strive to generate the anthropomorphic ship collision avoidance decision-making.In summary,this article has effectively improved the intelligent level of ship collision avoidance,reduced the risk of ship collision,and has practical significance for ensuring the safety of ships navigation.It provides an important theoretical basis for the application of data science in the maritime field,and provides research support for the application of artificial intelligence technology in ship driving direction.