基于自编码器与动态多元状态估计的核电厂主泵异常检测方法

Anomaly Detection Method for Main Pumps in Nuclear Power Plant Based on Autoencoder and Dynamic Multivariate State Estimation

摘要: 针对核电厂主泵运行数据具有高维、非线性和动态变化特征，提出一种融合自编码器与动态多元状态估计的主泵异常检测方法(AutoEncoder-Dynamic Multivariate State Estimation Technique, AE-DMSET)。该方法在动态多元状态估计(Dynamic Multivariate State Estimation Technique, DMSET)方法的基础上，引入自编码器学习主泵正常运行状态下多变量监测数据之间的非线性特征关系，并利用重构残差实现异常识别；同时，融合孤立森林(Isolation Forest, IF)、局部离群因子(Local Outlier Factor, LOF)和DBSCAN构建联合异常评分机制，对候选正常样本进行多角度筛选，并结合时间衰减因子实现记忆矩阵动态更新。基于某核电厂主泵历史运行数据开展实验，结果表明，AE-DMSET在动态工况下具有更稳定的状态重构能力和更可靠的异常检测性能，能够在保持较高异常检出能力的同时有效降低误报风险。AE-DMSET方法的精确率、召回率和F1分数分别达到0.978、0.943和0.960，能够提升核电厂主泵复杂动态工况下异常检测的准确性，具有一定工程应用价值。

Abstract: To address the high-dimensional, nonlinear, and dynamic characteristics of operational data from nuclear power plant main pumps, an anomaly detection method integrating an Autoencoder with Dynamic Multivariate State Estimation Technique (AE-DMSET) is proposed. Based on the Dynamic Multivariate State Estimation Technique (DMSET), the proposed method introduces an Autoencoder to learn the nonlinear feature relationships among multivariate monitoring data under normal operating conditions, and utilizes reconstruction residuals for anomaly identification. Meanwhile, a joint anomaly scoring mechanism combining Isolation Forest (IF), Local Outlier Factor (LOF), and DBSCAN is constructed to perform multi-perspective screening of candidate normal samples. A time decay factor is further incorporated to realize dynamic updating of the memory matrix. Experiments are conducted using historical operational data from a nuclear power plant main pump. The results demonstrate that AE-DMSET exhibits more stable state reconstruction capability and more reliable anomaly detection performance under dynamic operating conditions, while effectively reducing false alarm risks and maintaining high anomaly detection capability. The proposed AE-DMSET achieves a precision, recall, and F1-score of 0.978, 0.943, and 0.960, respectively, indicating that it can improve the accuracy of anomaly detection for nuclear power plant main pumps under complex dynamic operating conditions and has potential engineering application value.