Advances in probabilistic risk assessment (PRA) for decision-making under uncertainty

核工业的一个核心问题是需要在不确定的情况下做出高风险的决定。在英国，及时的例子包括决定延长当前核设施中组件的寿命，多个先进气体反应堆(AGR)工厂将在未来10年达到或超过其计划寿命。经过30多年的发展，风险评估任务的传统方法一直是应用基于概率风险分析的方法，描述故障情景的方法通常包括故障树、事件树、马尔可夫模型和最近的贝叶斯信念网络。在这些形式主义中捕获的数据通常使用纯粹的概率方法进行评估。然而，传统的PRA在表征某些形式的不确定性方面存在问题，例如，对潜在物理现象的知识有限。此外，基于模型的PRA的一个关键限制是在模型级别上执行概率不确定性量化(UQ)，从而将子结构/组件模型的预测集成到全局模型中，以便在不确定情况下做出系统级决策。本项目的目的是开发在贝叶斯概率风险评估框架内综合多种来源(数据和基于物理的模型)产生的不确定性的方法，使用图形模型作为组合框架，具体应用于安全关键和/或高价值结构的不确定性下的决策。这将涉及到学生学习从模型和数据中量化不确定性的最新技术，以及在拟议的框架内结合机器学习和数值模型验证的方法。该项目将从在LVV内进行的小规模测试发展到应用于真实世界的案例研究。这项工作构成了一个新的研究主题，它建立在监管团队目前在不确定性量化、模型验证和机器学习方面的兴趣之上。

A core problem in the nuclear industry arises from the need to make high-stakes decisions under uncertainty. Timely examples in the UK include decisions over the extension of life for components in current nuclear installations, with multiple Advanced Gas Reactor (AGR) plants due to reach or exceed their planned life over the next 10 years. The traditional approach to the risk assessment task, developed over more than 30 years, has been to apply methods based on probabilistic risk analysis, with approaches to characterise failure scenarios having typically included fault trees, event trees, Markov models and, more recently, Bayesian belief networks. The data captured within these formalisms has typically been evaluated using purely probabilistic methods. However, traditional PRA has issues in characterising certain forms of uncertainty, for example where knowledge of the underlying physical phenomena is limited. In addition, a key limitation in model-based PRA is in performing probabilistic uncertainty quantification (UQ) across model levels, and thus integrating the predictions of a substructure/component model into a global model to make system-level decisions under uncertainty. The aim of this project is to develop methods for combining uncertainties arising from multiple sources (data and physics-based models) within a Bayesian probabilistic risk assessment framework using graphical models as a combining framework, with the specific application being decision making under uncertainty for safety-critical and/or high-value structures. This will involve the student learning state-of-the-art techniques for quantifying uncertainty from both models and data; and for combining methods from machine learning and numerical model validation within the proposed framework. The project will build from small scale testing conducted within the LVV to application to real world case studies. The work constitutes a new research theme that builds on the current interests of the supervisory team in uncertainty quantification, model validation and machine learning.