Abstract: In the context of the national maritime power strategy and the Maritime Silk Road Initiative, vessels tend to be increasing in number and capacity. These factors have been combined to deteriorate the navigation environment and increase maritime navigation accidents. Due to the difficulty of search and rescue, these accidents can result in many casualties, huge economic losses and irreparable environmental damage. Therefore, it is crucial and timely to dig out risk factors of maritime navigation accidents.
According to the literature review, the existing researches on risk factors of maritime navigation accidents have the following shortcomings: (1)There is a lack comprehensive and objective consideration of the risk factors, especially detailed analysis of human factors, although many studies have shown that human factors are very important to maritime navigation accidents. (2)Maritime navigation accidents are with low probability. Due to the lack of public data, the accident sample size cannot meet most of the research models. In order to meet the demand of sample size, existing studies mainly obtain large-size samples by expanding the research area and data expansion. However, different areas have different risk factors of maritime navigation accidents, and data expansion cannot fully reflect the characteristics of data. Therefore, both area expansion and data expansion are not conducive to excavating the risk factors of maritime navigation accidents accurately. (3)The existing model can not be used to predict the probability of accidents, or directly used to analyse the impact of each risk factor on the accidents. (4)The existing studies mainly dig out risk factors of maritime navigation accidents by analysing the data of the vessels involved in navigation accidents, or compare the data of vessels in a certain type of accidents with other types of accidents. In fact, the vessels involved in the accidents do not necessarily have the characteristics of risk. To solve the first problem, this paper digs out the risk factors of maritime navigation accidents, including human, ship, management and environmental factors. In order to ensure that comprehensiveness of the risk factors, potential risk factors are mined from accident reports and literatures by text mining, and ensure the validity of the risk factors, these mined factors which are obviously different between the vessels liable for navigation accidents and the other vessels are screened out as risk factors by correlation analysis. For the second problem, this paper reduces the sample size requirement by feature optimization and proposing an improved SVM (support vector machine) model. Feature optimization can reduce the dimension of input variables, so as to reduce the sample size requirements and improve the accuracy of the model. As a supervised classification model, SVM requires less sample size and has been successfully used in the studies of traffic accidents. Existing studies show that the SVM model has certain advantages for mining risk factors of accidents, and has not been involved in the analysis of risk factors of maritime navigation accidents. For the third problem, the RFE (recursive feature elimination) algorithm is used to analyze the impact of the independent variable on the target variable. In view of the fourth problem, this paper digs out risk factors of maritime navigation accidents by comparing the differences between the ships responsible for the navigation accidents and other ships. In conclusion, based on feature optimization and SVM, a method is proposed to dig out risk factors of maritime navigation accidents and analyse their impact on navigation accidents. Firstly, comprehensive risk factors are mined, including human factors, management factors, environment factors and ship factors. Potential risk factors are mined from accident reports and literatures by text mining, and these factors which are obviously different between the vessels liable for navigation accidents and the other vessels are screened out as risk factors by correlation analysis. Then, we set the excavated risk factors as input features. The SVM model, whose parameter combination is optimized based on cross validation and swarm intelligence optimization, is proposed to identify whether the ship is risk-related or not. Finally, recursive feature elimination algorithm is applied to the SVM model to screen out and sort the crucial risk factors based on their impact on navigation accidents.
The data of water traffic accidents in Guangdong province from 2012 to 2020 are used to verify the applicability of the proposed method. The results show that the accuracy of the model (90.20%) is higher than that of the traditional SVM model (75%), which are helpful for maritime navigation accident prevention and control. On the one hand, by analyzing the accident rate of ships classified by the crucial risk factors, ships, environments, enterprises and operators with higher accident rate can be mined. Strengthening the control and guidance of these scenarios is conducive to improving the maritime navigation safety. On the other hand, the model is a strong predictor for whether the ship is risk-related or not. Therefore, effective measures to avoid navigation accidents could be gained by adjusting the state of the crucial risk factors.

Key words: traffic safety; navigation accidents; risk factor; SVM-RFE; optimization algorithm

摘要： 在“海运强国”战略和建设“海上丝绸之路”的大背景下,我国对船舶航行安全提出更高的要求。本文基于特征优选和支持向量机模型,挖掘出船舶航行事故的致因并分析各个因素对事故的影响程度。首先,通过文本挖掘和相关性分析对输入特征进行优选,筛选出航行事故责任船舶与其他船舶存在明显差异的因素作为航行事故致因。然后,构建基于SVM的船舶航行事故识别模型,并通过交叉验证及群体智能优化算法选择模型的最佳参数组合,得到最优的分类模型。最后,利用递归特征消除算法将上述致因对事故的影响程度进行排序和筛选,挖掘出事故的关键致因。通过广东省的水上交通事故实例验证模型的有效性,结果表明:本模型(正确度为90.1%)较传统单一的SVM模型(正确度为75.0%)具有更高的精度。研究结果可为减少船舶航行事故提供有效的科学建议。

关键词: 交通安全, 航行事故, 致因因素, SVM-RFE, 优化算法