海盗袭击严重威胁海运安全，准确的风险评价对航线规划与应急决策至关重要。为科学评价海盗袭击风险，本文提出了一种基于改进贝叶斯网络的海盗袭击风险评价模型。为解决传统模型对连续型数据易进行不合理划分的缺陷，本文引入了Fisher最优分割算法，提升了贝叶斯网络模型节点状态设定的准确性和科学性。在完成样本筛选和节点状态设定优化的基础上，采用期望最大化算法（EM）进行贝叶斯网络参数学习。同时，通过改进的树形结构学习算法有效降低了模型结构学习的复杂性。研究结果表明，与传统贝叶斯网络模型相比，本模型的预测准确率提高了5.44%，结构学习任务的复杂性降低了21.43%。此外，本文模型的预测准确率比随机森林模型和反向传播（BP）神经网络模型分别提升了1.23%和2.83%。本文模型可为海盗袭击风险的科学预测提供有力支持。

Pirate attacks pose a serious threat to maritime shipping security. An accurate risk assessment model is critical for route planning and emergency decision-making. To scientifically assess piracy risks, a pirate attack risk assessment model based on an improved Bayesian network was proposed in this study. The Fisher optimal segmentation algorithm was introduced to enhance the model’s node-state settings, classification accuracy and scientific validity. Based on the screened samples and node-state settings, the expectation-maximization (EM) algorithm was used for Bayesian network parameter learning. Meanwhile, an improved tree structure learning algorithm was adopted to reduce the complexity of model structure learning. The proposed model improves prediction accuracy by 5.44% compared with the traditional Bayesian networks while reducing structure learning complexity by 21.43%. In addition, compared with random forest and backpropagation (BP) neural network models, it achieves improvements of 1.23% and 2.83% in prediction accuracy, respectively. This study provides robust decision support for the scientific prediction of piracy risks.