Research On The Knowledge Base Of Ship Collision Avoidance Based On HSSVM And Convolutional Neural Networks

To make sure the safety of ship navigation,intelligent decision-making system for ship collision avoidance plays a very important role in automatic collision avoidance and it can help to prevent and even avoid ship collision accidents and reduce casualties.An intelligent decision-making system for ship collision avoidance can not only help promote safety of navigation and reduce the impact of human factors and human labor,but also is important for the development of the world shipping industry.It has become a main research directions and priority in the field of ship navigation safety.The knowledge base of ship collision avoidance is the key data base of ship intelligent collision avoidance decision.In previous studies,knowledge base of ship collision avoidance is mainly based on experience data storage,rule learning and knowledge reasoning.There are few researches in concern of trajectory knowledge base for collision avoidance,graph database storage and image feature fusion knowledge base,which is also the core content of this paper.How to complete the classification of ship collision avoidance automatically with the ship motion trajectory from historical AIS data is studied.It helps to establish a concise version of ship collision avoidance knowledge base is studied in this paper.Concerned with the massive new AIS data,the combined features of trajectory motion annotation and TCPA/DCPA annotation are input into the CNN-based ship trajectory similarity matching neural network to complete large-scale AIS data collision avoidance knowledge base construction.From the perspective of traffic safety,this paper does research on rules and methods for collision avoidance to provide data support for intelligent collision avoidance.The main research results of this paper are presented as follows:(1)AIS ship trajectory data related preprocessing technology is studied.To process the massive AIS trajectory data,the time stamp filtering is first carried out by using the time slice inverted index,and then the space range search based on the R tree index is used to complete the encountering data extraction from AIS data.To solve the problem of redundancy in AIS data,one improved dynamic Douglas Peucker algorithm is proposed to realize AIS data compression based on the study of the classic Douglas Peucker algorithm,taking into account the dynamic behavior points of acceleration,deceleration and turning data points.That is,the state stationary points are added to avoid losing key data points in the compression process.Regarding the problem of data completion,the Lagrange bilinear interpolation is used for the straight-line trajectory and Lagrange quadratic interpolation is used for the ship curve motion based on the analysis of the first-order and the second-order difference quotient.The interpolation helps to meet the requirements of two ship trajectory alignment.In concern of TCPA/DCPA calculation used in collision avoidance pattern recognition,a concise and efficient calculation method is defined.The space collision risk measure DCPA and time collision risk measure TCPA of the target ship are calculated by using the plane vector operation theory and relative motion geometry analysis to get the TCPA/DCPA labeling results.(2)One classification decision based on a hybrid method of weighted kNN and hyper-sphere SVM is put forward for ship meeting situations classification.In order to accelerate solving quadratic programming to make sure quick convergence of the hypersphere support vector machine algorithm,the samples with large contribution to the result hypersphere are selected firstly during training based on the sample weight determined by the center distance ratio.At the same time,multi threads technology and genetic algorithms(GA)+SMO are introduced to obtain the optimized parameters when training.For each test sample.its classification depends on the position between the point and each classified hypersphere.In particular,for the test data classification of complex intersection region.kNN is introduced.By selecting the nearest neighbor points between the training data set and the test sample in the intersection of hyper-spheres,category of the test sample is decided using the approximation measure,so as to improve the generalization ability of the algorithm.The experimental results show that the new algorithm has higher efficiency and better classification performance.According to the collision avoidance process,the trajectory sequence data of the whole collision avoidance process is extracted to construct the ship collision avoidance knowledge base.(3)CNN-based trajectory similarity matching algorithm of ship collision avoidance is put forward to construct the knowledge base of collision avoidance with massive new AIS data.The ship trajectories processed are spliced together after interpolation and time alignment,and the fusion features are used as the input data of trajectory similarity matching network based on CNN.The convolution kernel function is designed to realize the convolution operation based on knowledge base of ship collision avoidance obtained by HSSVM.By the convolution layer.pooling layer and activation function layer,the original input data is mapped to the hidden layer feature space.Then,the distributed feature representation is mapped to the sample label space through weighted summation of the full connection layer.In order to complete the construction of large-scale collision avoidance knowledge base,entropy loss is used to train deep learning network by constantly optimizing CNN network parameters.Finally,the classification matching results of ship collision avoidance are obtained.Based on this,the application of the ship collision avoidance image knowledge base and the ship collision avoidance feature knowledge base is given.Some limitations of the existing collision avoidance knowledge base are overcomed by the proposed methods and practical application examples of knowledge storage,knowledge representation and knowledge fusion are expanded.