为揭示水上事故调查报告中风险因素的样本内条件关联，以我国东南沿海某水域2015—2024年306条事故与险情记录的总体统计资料为基础，从其中173条事故记录中整理114份信息较完整的事故调查报告。首先依据事故经过、原因分析和责任认定等文本，对人员、管理、船舶、环境、事故类型及事故后果进行分类编码，形成114×21的节点状态矩阵；然后根据航行安全逻辑和事故文本资料人工构建贝叶斯网络结构，在固定网络结构下采用期望最大化（Expectation-Maximization, EM）算法迭代估计各节点条件概率表；在此基础上，结合节点状态频次、目标状态敏感性、欧几里得距离关联强度和替代结构对比，分析风险因素之间的条件依赖及结构设定对结果的影响。结果表明：事故类型节点中，碰撞、搁浅、沉没、火灾/爆炸和其他事故的概率分别约为38%、25%、10%、6%和21%；人员因素累计频次占50.35%，高于管理、船舶和环境因素；在所考察证据情景中，疏忽瞭望、未遵守航行规定、未采取安全航速、指挥与协调不当等节点与碰撞或操纵失误目标状态的概率变化关联较明显；替代结构会改变部分概率数值，但主要敏感节点总体保持一致。建模事故报告中的人员行为、管理状态、船舶状态与环境条件存在多因素条件关联，其概率分布受样本筛选和网络结构设定影响，不能外推为正常航行条件下的事故发生概率或变量之间的因果效应。

To characterize within-sample conditional associations among risk factors documented in maritime accident investigation reports, this study used aggregate statistics from 306 accident and incident records collected in a coastal water area in southeastern China during 2015-2024. From 173 accident records, 114 sufficiently documented investigation reports were compiled for modeling. Text concerning event processes, cause analyses, and responsibility determinations was coded into 21 nodes covering personnel, management, vessel, environment, accident type, and accident consequence, producing a 114×21 state matrix. A Bayesian Network structure was then constructed manually according to navigation-safety logic and accident-report evidence, and conditional probability tables of each node were estimated under the fixed structure through expectation-maximization (EM) iterations. Node-state frequencies, target-state sensitivity, Euclidean-distance-based association strength, and an alternative-structure comparison were used to analyze the conditional dependencies between risk factors and the impact of structural settings on the results. The results show that the probabilities of collision, grounding, sinking, fire/explosion, and other accidents in the accident-type node were approximately 38%, 25%, 10%, 6%, and 21%, respectively; the cumulative frequency of human factors accounted for 50.35%, which was higher than that of management, ship, and environmental factors. Deficient lookout, noncompliance with navigation rules, failure to maintain a safe speed, and inadequate command and coordination showed comparatively pronounced associations with probability changes in the collision or operational-error target states. Although the alternative structure changed some probability values, the principal sensitive nodes remained broadly consistent. Personnel behavior, management status, vessel status, and environmental conditions in the modeled accident reports exhibit multifactor conditional associations; their probability distributions are affected by sample selection and network-structure specification and cannot be extrapolated to accident occurrence probabilities under normal navigation conditions or to causal effects between variables.