Research on Anomaly Detection and Correction of Nuclear Power Plant Operation Data Based on GRU-MLP
-
摘要: 为改善核动力装置仪表与控制系统采集或存储的运行数据中出现的缺失、飘移和跳跃等数据质量问题，以便为运行数据分析和自动控制器提供更可靠的输入，提出了基于门控循环单元（GRU）与多层感知机（MLP）融合模型的核动力装置运行监测数据异常检测与校正的一体化方法。GRU-MLP融合模型提出了基于GRU模型的运行监测数据短时预测算法，为运行监测数据异常检测和校正提供参考依据，并且设计了实时校正机制，以提高GRU模型对包含异常运行数据的预测准确率。然后，利用MLP模型的非线性拟合能力优化“预测-异常检测”机制下使用的固定阈值为动态阈值，提高设计方法的异常检测准确率。最后，以某型核动力装置运行数据开展测试实验，从多角度分析并证明了所设计方法的准确性和可行性。
-
关键词:
- 核动力装置 /
- 数据预测 /
- 数据异常检测与校正 /
- 门控循环单元（GRU） /
- 多层感知机（MLP）
Abstract: In order to solve the data quality problems such as missing, drifting and jumping in the operation data collected or stored by the instrument & control system of nuclear power plant and provide more reliable input for operational data analysis and automatic controllers, a hybrid model based on Gated Recurrent Unit and Multilayer Perception (GRU-MLP) is proposed to detect and correct the abnormal operation monitoring parameters data of nuclear power plant. Firstly, the short-term prediction algorithm of operation data based on GRU model is studied to provide reference for anomaly detection and correction of operation data. Secondly, in order to improve the prediction accuracy of GRU model for the operation data containing anomalies, the real-time correction mechanism is used for optimization. Then, using the nonlinear fitting ability of MLP model, the fixed threshold used in the "prediction-anomaly detection" mechanism is optimized to the dynamic threshold, which improves the anomaly detection accuracy of the proposed method. Finally, the accuracy and feasibility of the proposed algorithm are verified through experiments based on the operation data of a certain nuclear power plant.-
Key words:
- Nuclear power plant /
- Data prediction /
- Data anomaly detection and correction /
- GRU /
- MLP
-
图 1 GRU-MLP融合模型算法策略示意图
Figure 1. Schematic Diagram of GRU-MLP Model Algorithm Strategy
图 2 随机跳变下数据异常检测准确率曲线
Figure 2. Accuracy Curve of Data Anomaly Detection under Random Jump
图 3 线性增长下数据异常检测准确率曲线
Figure 3. Accuracy Curve of Data Anomaly Detection Under Linear Growth
图 4 未实时校正的GRU模型预测效果
Figure 4. Prediction Effect of GRU Model without Real-Time Correction
表 1 异常数据检测实验设置
Table 1. Settings of Data Anomaly Detection Experiments
实验编号 异常类型 参数 数据长度/s 工况 异常开始时间/s 异常变化形式 1 随机跳变 稳压器压力 7000 15%稳态 100随机点 最大为均值的10% 2 固定值漂移 稳压器水位 5000 15%稳态 2500 均值的5% 3 增长性漂移 线性漂移 蒸汽发生器压力 30%稳态 2500 Δy=0.001×(t−2500)+0.00002 4 多项式漂移 一回路平均温度 15%稳态 2500 Δy=7×10−6×(t−2500)2+0.35 5 对数漂移 稳压器压力 3000 50%稳态 1800 Δy=ln (t−1800) Δy —异常幅值随时间变化的函数 
下载: 导出CSV
表 2 异常数据检测平均准确率
Table 2. Average Accuracy of Data Anomaly Detection
实验编号 1 2 3 4 5 平均准确率 训练 0.9872 0.9945 0.8962 0.9750 0.9897 测试 0.9734 0.9920 0.8873 0.9462 0.9804
下载: 导出CSV
表 3 泛化能力验证实验设计
Table 3. Settings of Generalization Ability Experiments
实验编号 工况 故障情况 参数 数据长度/s 1 15%额定功率/主泵低速 稳态运行 左回路冷却剂流量 500 2 15%额定功率/主泵高速 稳态运行 左回路冷却剂流量 500 3 15%额定功率/主泵高速 稳态运行 稳压器水位 500 4 50%额定功率/主泵高速 稳态运行 稳压器水位 500 5 50%额定功率/主泵高速 左回路热端破口（当量直径0.004 m） 稳压器水位 500
下载: 导出CSV
表 5 异常数据检测算法比较实验设置
Table 5. Settings of Experiment on Abnormal Data Detection Algorithms
异常类型 参数 数据长度/s 异常开始时间/s 异常点数 随机跳变 稳压器压力 600 随机 100 固定偏移 500 400 100
下载: 导出CSV
表 6 状态判断正确率
Table 6. Accuracy of State Judgment
异常类型 GRU-MLP WinCen ARIMA 随机跳变 0.994 0.69 0.7633 固定偏移 0.9833 0.788 0.806
下载: 导出CSV
表 7 异常数据检出率
Table 7. Anomaly Detection Rate
异常类型 GRU-MLP WinCen ARIMA 随机跳变 1 0.16 0.8 固定偏移 1 0.04 0
下载: 导出CSV
-
[1] WANG X, WANG C. Time series data cleaning: a survey[J]. IEEE Access, 2020, 8: 1866-1881. doi: 10.1109/ACCESS.2019.2962152 [2] BASU S, MECKESHEIMER M. Automatic outlier detection for time series: an application to sensor data[J]. Knowledge and Information Systems, 2007, 11(2): 137-154. doi: 10.1007/s10115-006-0026-6 [3] KEOGH E, LIN J, FU A W. HOT SAX: efficiently finding the most unusual time series subsequence[C]//Proceedings of the Fifth IEEE International Conference on Data Mining. Houston: IEEE, 2005: 226-233. [4] WEI L, KEOGH E, XI X P. SAXually explicit images: finding unusual shapes[C]//Proceedings of the Sixth International Conference on Data Mining. Hong Kong, China: IEEE, 2006: 711-720. [5] LI D, CHEN D C, JIN B H, et al. MAD-GAN: multivariate anomaly detection for time series data with generative adversarial networks[C]//Proceedings of the 28th International Conference on Artificial Neural Networks. Munich: Springer, 2019: 703-716. [6] 孙英杰,彭敏俊. 基于MSET和SPRT的核动力装置异常状态监测技术研究[J]. 核动力工程,2015, 36(3): 57-61. [7] 赵梦薇,陈智,廖龙涛,等. 基于智能预测的反应堆功率调节研究[J]. 核动力工程,2019, 40(4): 166-171. [8] 蒋波涛,黄新波,WESLEY H J,等. 基于ν-支持向量机的事故工况下反应堆功率预测[J]. 核动力工程,2019, 40(6): 105-108. [9] 余刃,孔劲松,骆德生,等. 基于运行数据分析的核动力装置异常运行状态监测技术研究[J]. 核动力工程,2013, 34(6): 156-160. doi: 10.3969/j.issn.0258-0926.2013.06.037 [10] SILVA L O, ZÁRATE L E. A brief review of the main approaches for treatment of missing data[J]. Intelligent Data Analysis, 2014, 18(6): 1177-1198. doi: 10.3233/IDA-140690 [11] BAYER J, WIERSTRA D, TOGELIUS J, et al. Evolving memory cell structures for sequence learning[C]//Proceedings of the 19th International Conference on Artificial Neural Networks. Limassol: Springer, 2009: 755-764. [12] CHO K, VAN MERRIËNBOER B, GULCEHRE C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing. Doha, Qatar: Association for Computational Linguistics, 2014: 1724-1734. [13] ZHANG Z C, ZHOU D, ZHANG R F, et al. Medical entity relationship recognition based on bidirectional GRU and attentional mechanism[J]. Computer Engineering, 2020, 46(6): 296-302. -
计量
- 文章访问数: 169
- HTML全文浏览量: 23
- PDF下载量: 78
- 被引次数: 0
