The Study on Multivariate Time Series Anomaly Detection via Convolutional Variational Autoencoder

Monitoring is essential for managing and maintaining industrial equipment by identifying abnormal areas in multivariate time series. This work presents CNN-VAEAT, a new technique for identifying anomalies in multivariate time series data from complex systems. The approach uses an upgraded Variational Autoencoder (VAE) algorithm with CNNs to reconstruct the data and calculate an anomaly score. To select anomaly detection threshold, an adaptive threshold approach based on extreme value distribution theory is used. The approach reconstructs a normal multivariate time series model using a convolutional variational autoencoder. It then derives the system’s multivariate time series state index from the new signal. Finally, it uses the extended extreme value distribution theory to discover anomalous conditions early. Studies on the SCADA data set of real wind turbine faults show that the CNN-VAE-AT algorithm outperforms existing wind turbine fault omen detection techniques.