基于特征组合分析的主泵异常检测方法

为了解决传统阈值法在核电站主泵状态数据异常检测中的误判、实时性差等问题,提出一种基于单维状态数据特征分析和多维状态数据特征分析相结合的方法。对于单维状态参量,使用AR(auto regressive)模型拟合获得模型参数,再结合SOM(self organizing maps)神经网络的量化结果得到单维状态参量随时间变化的过渡概率序列;对于多维状态参量,使用OPTICS(ordering points to identify the clustering structure)算法聚类生成不同的模式组;然后根据两类特征提取结果综合分析,得到异常检测模型;最后将检测模型应用于主泵状态数据异常检测,并与其他方法进行比较。实验结果表明此模型在准确性、实时性上更具优势。

Anomaly Detection Method of Main Pump based on Feature Combination Analysis

In order to solve the problem of misjudgment and poor real-time performance of traditional threshold method in anomaly detection of main pump state data in nuclear power plant, a method based on single-dimensional state data and multi-dimensional state data feature analysis was proposed. For the feature extraction of single-dimensional state data,the AR model was used to fit the single-dimension state data to obtain the model parameters, and then the transition probability sequence of the single-dimension state data was obtained by combining the quantization results of SOM neural network. For the feature extraction of multi-dimensional state data, the OPTICS algorithm was used to cluster to generate different pattern groups. Based on the analysis of the two kinds of feature extraction results, an anomaly detection model was obtained. Finally, the detection model was applied to the anomaly detection of the main pump state data and compared with other methods. The experimental results show that the model has advantages in accuracy and real-time.