张大志, 罗骁域, 郑胜. 基于POT的多元统计过程核电数据异常检测方法[J]. 南方能源建设. DOI: 10.16516/j.ceec.2024-099
引用本文: 张大志, 罗骁域, 郑胜. 基于POT的多元统计过程核电数据异常检测方法[J]. 南方能源建设. DOI: 10.16516/j.ceec.2024-099
ZHANG Dazhi, LUO Xiaoyu, ZHENG Sheng. An Anomaly Detection Method for Multivariate Statistical Process Based on POT[J]. Southern Energy Construction. DOI: 10.16516/j.ceec.2024-099
Citation: ZHANG Dazhi, LUO Xiaoyu, ZHENG Sheng. An Anomaly Detection Method for Multivariate Statistical Process Based on POT[J]. Southern Energy Construction. DOI: 10.16516/j.ceec.2024-099

基于POT的多元统计过程核电数据异常检测方法

An Anomaly Detection Method for Multivariate Statistical Process Based on POT

  • 摘要:
      目的  核电设备的安全运行对核电厂至关重要,发生事故所带来的损失是不可估量的。因此,对核电设备进行有效的异常检测十分必要。针对固定阈值和人为检测方法的局限性,这些方法难以适应时序数据的动态变化,文章提出一种基于POT的多元统计过程的异常检测方法。
      方法  文章采用主成分分析方法构建异常检测模型,将模型的SPE统计量作为POT算法的初始阈值,然后将超过初始阈值的部分进行广义帕累托分布拟合,从而确定最终的动态阈值。当异常分数超过最终阈值则发出异常警告。通过将多元统计过程控制和极值理论相结合,该方法利用多元统计过程控制快速发现核电厂运行数据中的异常情况,并结合极值理论对极端事件的建模与分析来提高异常检测的灵敏度和可靠性,能够快速发现核电厂高维运行数据中存在的异常情况。
      结果  在仿真实验结果中,文章提出的方法相较于常规的多元统计方法和POT方法,具有更高的准确率、召回率。在核电厂不同设备上的实际运行数据的实验中,证明了该方法在异常检测上的有效性。
      结论  将多元统计过程控制和极值理论结合,提出的异常检测方法不仅能检测到由数据相互关系改变引起的异常,而且能利用POT方法确定最终阈值避免传统多元统计过程控制中出现的误检。该方法能处理核电厂高维时序运行数据,提高异常发现的效率,确保了核电厂安全高效地运行从而提高核电厂的经济效益。

     

    Abstract:
      Introduction  The safe operation of nuclear power equipment is crucial for nuclear power plants (NPPs), and the losses caused by accidents are immeasurable. Therefore, effective anomaly detection for nuclear power equipment is necessary. Considering the limitations of fixed thresholds and manual detection methods, which are difficult to adapt to the dynamic changes in time series data, this paper proposes an anomaly detection method based on POT for multivariate statistical processes.
      Method  This paper adopted PCA to construct an anomaly detection model, where the SPE statistic of the model served as the initial threshold for the POT algorithm. Subsequently, the portion exceeding the initial threshold was fitted with a generalized Pareto distribution to determine the final dynamic threshold. An anomaly warning was issued when the anomaly score exceeded the final threshold. By combining multivariate statistical process control (MSPC) with extreme value theory (EVT), this method used MSPC to discover anomalies in the operating data of NPPs quickly and improved the sensitivity and reliability of anomaly detection by modeling and analyzing extreme events, so that it can quickly detect anomalies in high-dimensional operating data of NPPs.
      Result  In the simulation experiment results, the proposed method has a higher accuracy and recall rate than conventional multivariate statistical and POT methods. In experiments with actual operating data from different equipment in NPPs, the method's effectiveness in anomaly detection has been demonstrated.
      Conclusion  By combining MPSC with EVT, the anomaly detection method proposed in this paper can not only detect anomalies caused by changes in data relationships but also avoid false detection in traditional MSPC by determining the final threshold using the POT method. This method can handle high-dimensional time series operating data of NPPs, improve the efficiency of anomaly detection, ensure the safe and efficient operation of NPPs, and improve their economic benefits.

     

/

返回文章
返回