DUCG动态立体因果图的构建和推理方法及其实验验证研究

Fully considering the development trend of the fault diagnosis technology,difficulties and application demands, based on the present Dynamic Uncertain Causality Graph (DUCG) framework and knowing its incapability, this study will focus on solving, in particularly in the cases of negative feedback, the probabilistic reasoning and logic interpretation to the inference results. In the new theory, the inference engine will autimatically construct the specific cubic-DUCG online, allow any causality penetration among different time slices, and make dynamic uncetain inference and prediction. This mode of directly constructing dynamic model to deal with dynamic problem changes the present popular mode of treating a dynamic problem as a number of static problems, which is commonly used in the international academic community.The suggested theory is a new thought, and is an important development to the present DUCG theory originally presented in China, particularly useful to solve large and complex dynamic intelligent fault diagnosis problems，interprete the inference results and make predictions. The research group will study the rules how to construct the cubic DUCG according to the time sequence, the causality influence delay and overlap, multi-parameter inference, the treatment to the spurious signals and imperfect knowledge base, fuzzy or uncertain evidence, dynamic negative feedback, efficient calculation, graphical showing of the inference results, development of the software system enbodied with the new theory, hand calculation and doing experiments in a nuclear power plant simulator to validate the correctness, usability and efficiency of the new theory.

本研究充分考虑智能故障诊断技术的发展趋势、困难和应用需求，以现有动态不确定因果图DUCG理论框架为基础，并针对其不足，重点解决存在负反馈情况下的概率推理和逻辑解释问题。在新理论中，推理机将自动在线构造特定情况下的立体DUCG图，允许因果关系穿越任意时间片，并据此进行动态不确定推理和预测。这种直接用动态模型来处理动态问题的方式改变了过去学术界普遍采用的将动态问题静态化处理的基本模式，是一种全新的思路，也是对我国原创的DUCG理论体系的重要扩展，特别适用于解决大型复杂系统动态工况下的故障智能诊断、推理结果解释和预测问题。课题组将研究如何按照时序构建立体因果图的规则、因果影响的延滞和叠加、多参数组合推理、虚假信号和有错知识库处理、模糊证据、动态负反馈，高效计算，推理结果可视化解释等关键技术，开发基于新理论的软件系统，手算和在核电模拟机上做在线故障实验，以验证新理论的正确性、实用性和效率。

本课题面向核电站等大型复杂系统的动态、实时和高可靠性的运维决策需求，以DUCG动态不确定因果图理论为基础，针对现有理论方法的不足，研究建立可用于负反馈等复杂工况的工程系统动态工况下的故障智能监测、诊断与发展预测的新理论模型——立体DUCG（Cubic DUCG），包括依据原始DUCG根据在线收集到的动态证据自动构建反映时序工况的立体DUCG因果图，并基于立体因果图进行实时动态的高效准确的动态逻辑推理、立体DUCG图的动态演化和推理结果的图形生成和解释等关键技术和方法，重点解决在存在动态负反馈情况下的立体DUCG建模和概率推理问题，开发基于新理论的软件系统，在通过手算数学算例验证后，在宁德核电模拟机系统平台上进行在线故障诊断实验，通过24个典型故障实验，验证了立体DUCG理论方法的正确性、准确性（正确率100%）和算法的高效性性（能够在普通笔记本电脑上每秒钟完成一次推理）。这一研究成果采用动态的方法解决动态的复杂问题，为国际上首创，为大型复杂系统在线安全运维提供了有效的解决方法，可大幅提高系统的可靠性和可用度、显著提升核电站的安全性和经济效益，具有广泛的工程应用价值。目前已获得国家环保部连续两年的进一步资助。