A Nuclear Reactor Accident Diagnosis Technology Integrating Expert Knowledge and Machine Learning Algorithms
-
摘要: 核反应堆事故诊断是事故处理过程中最为重要的一步，诊断结果直接决定了事故的处置策略。本文提出了一种融合专家知识与机器学习算法的核反应堆事故诊断方法，该方法在已有的成熟的专家知识基础上，引入机器学习诊断算法，实现两类方法优势的叠加和劣势的互补。在专家知识诊断方面，采用基于征兆导向的事故诊断方法，形成征兆导向专家知识库，并形成基于专家知识的事故诊断功能模块；在基于机器学习算法的事故诊断方面，采用极端梯度提升（XGBoost）算法、线性支持向量机（SVM）算法、深度前馈网络（DFN）以及长短期记忆（LSTM）算法建立了事故诊断模型，并利用投票机制算法对各类算法进行融合，形成了机器学习智能诊断模块。在此基础上，提出了以专家知识诊断模型为主，机器学习智能诊断为辅的诊断模型，并利用“华龙一号”蒸汽发生器传热管破裂（SGTR）事故进行了验证，证明了该方法的合理性。Abstract: The nuclear reactor accident diagnosis is the most important step in the accident handling process, and the diagnosis result directly determines the accident handling strategy. In this paper, a nuclear reactor accident diagnosis method is proposed, which combines expert knowledge and machine learning algorithms. Based on the existing mature expert knowledge, this method introduces machine learning diagnosis algorithms to realize the superposition of advantages and the complementarity of disadvantages of the two methods. In terms of expert knowledge diagnosis, the symptom-oriented accident diagnosis method is adopted to form a symptom-oriented expert knowledge base and an accident diagnosis function module based on expert knowledge; In terms of accident diagnosis based on machine learning algorithm, Extreme Gradient Boosting (XGBoost), Linear Support Vector Machines (SVM), Deep Feedforward Networks (DFN), and Long Short-Term Memory (LSTM) are used to establish the accident diagnosis model, and the voting mechanism algorithm is used to fuse all kinds of algorithms to form the machine learning intelligent diagnosis module. On this basis, this paper puts forward a diagnosis model based on expert knowledge, supplemented by machine learning intelligent diagnosis, and the verification is carried out using the Steam Generator Tube Rupture (SGTR) accident of HPR1000. The results prove the rationality of the method.
-
Key words:
- Expert knowledge /
- Machine learning /
- Accident diagnosis /
- Model fusion
-
图 1 征兆导向专家知识库建立步骤
Figure 1. Steps for Establishing a Symptom-oriented Expert Knowledge Base
图 2 线性可分情况下的最优分类超平面
Figure 2. Optimal Classification Hyperplane in Cases of Linear Separability
图 4 LSTM节点的计算流程图
c—单元状态； h—节点输出；f—遗忘门； i—输入门；o—输出门；w—权重矩阵；tanh—双曲正切函数；φ—激活函数；t—当前时刻x输入信号；t-1—上一时刻 x输入信号
Figure 4. Computational Flowchart of LSTM
表 1 各类算法事故诊断结果
Table 1. Accident Diagnosis Results of Various Algorithms
事故类型 算法 P R F1值 冷管段LOCA XGBoost 0.974 0.973 0.974 SVM 0.920 0.950 0.930 DFN 0.740 0.830 0.780 LSTM 0.990 0.970 0.980 蒸汽发生器二次泄漏 XGBoost 0.995 0.995 0.995 SVM 1.000 0.990 1.000 DFN 0.920 0.940 0.930 LSTM 1.000 1.000 1.000 满功率下的SGTR事故 XGBoost 0.974 0.973 0.972 SVM 0.990 0.980 0.990 DFN 0.940 0.910 0.930 LSTM 1.000 1.000 1.000 
下载: 导出CSV
-
[1] 王晓龙,蔡琦,张晓奇,等. 基于多分类SVM的船用核动力装置主回路系统破口特征诊断技术研究[J]. 原子能科学技术,2014, 48(3): 462-468. doi: 10.7538/yzk.2014.48.03.0462 [2] 佘兢克,薛时雨,孙培伟,等. 基于深度学习的核电站事故预测及故障诊断方法[J]. 仪器仪表用户,2019, 26(12): 39-44. doi: 10.3969/j.issn.1671-1041.2019.12.010 [3] 钱虹,骆建波,金蔚霄,等. 报警触发式蒸汽发生器传热管破裂事故诊断专家系统的研究[J]. 核动力工程,2015, 36(1): 98-103. [4] 潘进,丁强,江爱朋,等. 基于XGBoost的冷水机组不平衡数据故障诊断[J]. 机械强度,2021, 43(1): 27-33. [5] 毛伟,余刃,陆古兵. 基于LS-SVM方法的某核电站主泵故障诊断[J]. 海军工程大学学报,2012, 24(5): 82-85. doi: 10.7495/j.issn.1009-3486.2012.05.017 [6] 许立武,李开成,肖贤贵,等. 基于深度前馈网络的电能质量复合扰动识别[J]. 电测与仪表,2020, 57(1): 62-69,130. [7] HOCHREITER S, SCHMIDHUBER J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735-1780. doi: 10.1162/neco.1997.9.8.1735 [8] YANG Q H. Research on SVM remote sensing image classification based on parallelization[J]. Journal of Physics: Conference Series, 2021, 1852: 032009. doi: 10.1088/1742-6596/1852/3/032009 [9] 王丽华,谢阳阳,周子贤,等. 基于卷积神经网络的异步电机故障诊断[J]. 振动、测试与诊断,2017, 37(6): 1208-1215. -
计量
- 文章访问数: 24
- HTML全文浏览量: 10
- PDF下载量: 12
- 被引次数: 0
