基于CAE和OCSVM的地铁道岔异常检测

Anomaly Detection of Metro Railway Turnout Based on CAE and OCSVM

摘要: 为提高地铁道岔设备异常检测的准确率，提出了基于卷积自编码器与单类支持向量机的异常检测框架。以ZDJ9型交流电动转辙机为研究对象，使用现场运行过程中采集到的三相电流数据。首先，在数据预处理阶段，为保留原始数据的时频信息，本文利用短时傅里叶变换将原始一维时间序列数据转化为时频矩阵。同时，为综合利用多域信息，将每一相电流数据所对应的时频图堆叠，构成三维特征信息。然后，采用卷积自编码器对预处理后的图像数据进行进一步的降维与特征提取。最后，将提取出的特征输入到单类支持向量机进行异常检测模型训练。实验结果表明，相较于直接使用原始一维时间序列进行异常检测，结合时频与多域信息的模型具有更高的准确率以及F1-Score。

Abstract: To improve the accuracy of metro turnout equipment anomaly detection, an anomaly detection framework based on the convolutional autoencoder and one-class support vector machine is proposed herein. Taking ZDJ9 AC electric switch machine as the research object, the three-phase current data collected during field operation are used. First, in the data preprocessing stage, to preserve the time-frequency information of the original data, the short time Fourier transform is used to convert the original one-dimensional time series data into a time-frequency matrix in this paper. At the same time, in order to comprehensively utilize multi-domain information, the time-frequency diagrams corresponding to each phase of current data are stacked to form three-dimensional feature information. Then, a convolutional autoencoder is used to perform further dimension reduction and feature extraction of the preprocessed image data. Finally, the extracted features are input to a one-class support vector machine for anomaly detection model training. The experimental results show that the model combining time-frequency and multi-domain information has higher accuracy and F1-Score than directly using the original one-dimensional time series for anomaly detection.