Unsupervised anomaly detection with a temporal continuation, confidence-aware VAE-GAN

We propose an unsupervised approach to anomaly detection in data with a temporal dimension. We adapt the VAE-GAN architecture to learn the proxy task of temporal sequence continuation. Rather than reconstructing the input, our variational decoder decodes to a forecast of the future sequence. In order to separate structural uncertainty (which our model can reconstruct by fitting to observed data) from stochastic uncertainty (which it cannot) we introduce an additional decoder that outputs the pointwise confidence of the prediction, after the optimal latent-variable has been found. We can use this for zero-shot anomaly detection, separating anomalies from stochastic variation that cannot be modelled, without any examples. This is important for domains in which anomalies are so rare that it is not possible or meaningful to train a supervised model. As an example of such a domain, we introduce a new dataset comprising linescan imagery of railway lines which we use to illustrate our methods. We also achieve state-of-the-art performance on the ECG5000 and MIT-BIH time series anomaly detection datasets. We make an implementation of our method available at https://github.com/YorkXingZeyu/ECG-VAEGAN-Project.