基于多尺度全卷积神经网络的核电主泵状态异常检测方法

Reactor coolant pump status anomaly detection method based on multi-scale fully convolutional networks

为解决目前主泵状态异常检测复杂数据建模难、早期异常检出难和异常程度评估难的问题,提出一种基于多尺度全卷积神经网络的编解码器网络结构异常检测方法MSFCNAD(multi-scale FCN-based anomaly detection)。在考虑主泵状态多变量时序特征的基础上,利用全卷积神经网络编解码进行像素级训练,精准定位核电主泵状态数据的异常范围;同时兼顾主泵状态异常时段特征,提取核电主泵状态的多尺度特征矩阵,通过不同尺度的特征矩阵所检测到的异常范围分级判断其异常程度。在此基础上,利用真实核电主泵数据进行实验,对比自回归滑动平均模型(ARMA)、双向长短期记忆网络(BiLSTM)、全卷积神经网络(FCN)、自动编码器(AEs)等多个模型的分类效果。结果显示,MSFCNAD模型的召回率、F1分数均优于文中列举的模型,分别为78.44%、80.30%,优于其他模型中最高的77.53%、69.74%,表明MSFCNAD模型比其他异常检测方法的性能更好,同时能够通过异常程度判断异常的严重性,在故障发生前优先维护,保障主泵正常运行。

The large number of types of reactor coolant pump condition sensors leads to three difficulties in current main pump condition anomaly detection:difficulty in modeling complex condition data,difficulty in detecting early abnormalities,and difficulty in assessing the degree of abnormality.The study of rapidly developing deep learning techniques such as neural networks can provide new ideas to solve these problems.To this end,a codec network structure anomaly detection method MSFCNAD(multi-scale FCN-based anomaly detection)based on multi-scale fully convolutional neural networks is proposed.Based on the multivariate temporal characteristics of the main pump state,this method uses the full convolutional neural network codec for pixel-level training to precisely locate the abnormal range of reactor coolant pump state data.At the same time,taking into account the main pump state abnormal time characteristics,the multi-scale feature matrix of reactor coolant pump state is extracted,and the abnormal extent is judged by the graded abnormal range detected by the feature matrix of different scales.On this basis,experiments are conducted using real nuclear main pump data to compare the classification effects of several models such as ARMA,BiLSTM,FCN and AEs.The results show that the MSFCNAD model outperforms the models listed in the paper in terms of recall and F1 score,which are 78.44% and 80.30%,respectively,better than the highest 77.53% and 69.74% of the other models.The experimental results show that this method has better performance compared with other anomaly detection methods,and it can also judge the severity of anomalies by the degree of anomalies and prioritize maintenance processing.