Anomaly detection for multivariate times series through the multi-scale convolutional recurrent variational autoencoder

To realize the anomaly detection for industrial multi-sensor data, we develop a novel multi-scale convolutional recurrent variational autoencoder (MSCRVAE) model. It is a hybrid of convolutional autoencoder and convolutional long short-term memory with variational autoencoder (ConvLSTM-VAE). The ConvLSTM-VAE part helps the MSCRVAE model not only capture the spatial and temporal dependence of data but also learn robust high-level representations by estimating the distributions of latent variables. Moreover, a typical loss function that combines the characteristics of both autoencoder and variational autoencoder is designed for the MSCRVAE model. The experimental results on three public data illustrate the superiority of the MSCRVAE model in anomaly detection with the best average F1-score up to 0.90. The interquartile range of the boxplots on the public data also proves the robustness of the MSCRVAE model. We also illustrate the role of the two components in the loss function by exploring their changes during training. What is more, on the private data, the MSCRVAE model also performs well and the residual matrices provide reasonable interpretations to the anomaly detection results.