安全与环境学报

为解决传统静态评估方法难以适应受限空间作业风险动态演化特性的问题，本文提出一种融合离散贝叶斯网络与非线性风险计算公式的实时评估模型。通过构建涵盖“人–机–管–环”的多因素贝叶斯网络，以动态节点表征气体浓度与人员行为等关键参数，量化事故概率；并引入Sigmoid函数对传统风险公式进行非线性重构，精准刻画环境动态节点异常在临界阈值附近的风险突变与非线性跃迁效应。基于贝叶斯网络推理引擎与多源数据同步机制，实现事故发生概率的在线更新与风险值的实时计算。通过受限空间作业系统构建和试验验证可知，该模型在评估准确度与预警灵敏度方面均优于传统方法，尤其对“低概率–高后果”事件的风险的识别能力显著增强，为受限空间作业安全管理的动态风险预控提供有效评估工具。

To address the critical deficiency of conventional static risk assessment methodologies in adapting to the inherently dynamic and evolutionary nature of hazards during confined space operations， this paper proposes a real-time assessment framework that seamlessly integrates a discrete Bayesian network with a nonlinear risk quantification algorithm. Based on a comprehensive investigation of frequent accident precursors in such environments， a multi-factor Bayesian network is constructed， encompassing the four dimensions of human， machine， management， and environmental factors. Within this structure， dynamic nodes specifically characterizing gas volume concentration and personnel unsafe behaviors are defined to capture the primary drivers of risk evolution. The temporal progression of accident probabilities is quantified in real-time via the PySmile inference engine， which computes conditional probabilities based on the dynamic node states. Subsequently， the conventional linear risk formulation is fundamentally improved by introducing a Sigmoid-based transformation function that establishes a sophisticated nonlinear mapping between the absolute deviation of gas concentrations exceeding critical limits and the resultant risk magnitude. By analyzing the differential contributions of abnormal environmental parameters and unsafe acts to the overall consequence severity， the model precisely captures the mutation characteristics and nonlinear transition effects that manifest when environmental parameters approach critical thresholds， thereby enabling the dynamic correction of risk values. Furthermore， to mitigate transient noise interference inherent in high-frequency sensor data， a unified time-benchmarking strategy coupled with a sliding window decision mechanism is engineered within the anomaly detection module， ensuring a robust balance between sampling frequency and node state updating. An experimental system was established based on a typical storage tank confined space scenario， architected into four hierarchical layers comprising physical perception， data acquisition and fusion， core computation， and decision support. Deploying networked gas sensors and video surveillance equipment in authentic operational environments， the system facilitates real-time gas monitoring and algorithmic recognition of critical personnel behaviors. Validation across six distinct operational scenarios demonstrates that the proposed model significantly outperforms traditional methods in both assessment accuracy and early warning sensitivity. Notably， under dynamically coupled multi-factor scenarios， the model accurately identified first-level major risks， whereas conventional models consistently underestimated these events as general risks. The results indicate that this model effectively compensates for the evaluation blind spots of conventional methods at critical risk transition points， providing a robust and effective tool for the dynamic pre-control of safety management in confined space operations.