Important Variables Identification and Proactive Evaluation of Real-time Ship Traffic Sailing Risk in Waterway

In order to further improve accident prediction accuracy of real-time ship traffic in waterway, based on ship detector data and traffic accident data collected on two downstream waterways of Yangtze River, important variables were sifted with random forest (RF) model from the initial data of waterway status within 20-40min before the traffic accident, then new Bayesian network(BN)model was established with 4 most important variables combined with Gaussian mixture model (GMM) and maximum expectation(EM) algorithm. Compared with BN model previous studied built with direct initial data, the new models complexity is not only reduced and its prediction effect is improved, with the accident prediction correct rate of 81.29%. The transfer-abilities of two models were also compared, the results show that though accident prediction correct rate of new model is decreased a little, its value is still higher than old model, moreover, new model is much better than old model in terms of non-accident and overall prediction correct rate and their rates are 94.63% and 91.86% respectively.