Unsupervised Auto-Encoder-Based Ensemble Algorithm for Anomaly Detection

These days, multidimensional time series data are universally collected and analyzed in various practical situations, such as intelligence factories and wearable devices. Anomaly detection in multivariable time series refer to deviating abnormal states that are different from normal behaviors and locating their causes. However, building such a system is challenging because the normal behavior of a system depends not only on the time dependencies in each time series, but on the interrelationships lying behind different time series pairs. Besides, the system is required to be robust to noise and provides operators with different levels of anomaly scores based on the severity of different events. This paper proposes an Unsupervised Auto-Encoder-Based Ensemble Algorithm (UAEE) for anomaly detection in multivariable time series data. Specifically, UAEE constructs sensor difference maps from the data of different time steps, and encoders as the input to the encoders based on convolutional neural network and pointwise convolutional network to capture both spatiotemporal patterns. Finally, the convolution decoder is used to reconstruct the map based on the feature map of spatiotemporal information, and the residual feature matrix is used for anomaly detection and diagnosis. Extensive empirical studies based on real test bed datasets have shown that Unsupervised Auto-Encoder-Based Ensemble Algorithm (UAEE) can outperform the other baseline approaches.