A MMEM-BN–Based Analyzing Framework for Causal Analysis of Ship Collisions

Exploring the causes and evolution of ship collisions plays an important role in accident prevention and risk control. This paper proposes a hybrid analyzing framework based on man–machine–environment–management (MMEM) and a Bayesian network (BN), and applied the analyzing framework to causal analysis of ship collisions in Zhejiang coastal waters based on 107 cases of ship collision accidents. The MMEM frame is utilized to guide influencing factors (IFs) and results of IF identification and causal chain analysis are used as basis for structuring the BN model, using 1 and 0 to indicate whether an IF occurs or not, respectively, in regard to an accident. This marking method was used to deal with each of 107 accident cases in order to form set of sample data, of which 97 cases were used as training cases, 10 cases were used as testing cases. The model was trained by using the set of sample data and the expectation maximization (EM) algorithm to obtain the conditional probability tables (CPTs). A 20-node BN model with the ability to predict the probability of occurrence of ship collision was established. The BN model was verified through sensitivity analysis–based validation, k-folds cross-validation, and testing cases–based validation. The validation results show the correctness and usability of the model. Through backward reasoning–based analysis, maximum likelihood cause chain analysis, and sensitivity analysis, it was found that human error (H) is the main IF resulting in ship collisions; the causal chain that maximizes the likelihood of an accident occurring is H1 (Improper lookout) → H4 (Underestimation of collision) → H7 (Failure to take effective collision avoidance measures) → H (Human error) → C (Ship collision). H1, H7, and H9 (Improper emergency handling) have relatively high sensitivity and greater impact on collision accidents. The results show that the proposed MMEM-BN–based analyzing framework is applicable. The analyzing framework and results of causal analysis will provide theoretical and practical supports for exploring origin of accidents, revealing evolutionary mechanism of accidents and for taking targeted risk control measures.