Anomaly Detection in Multivariate Time Series using Graph Convolutional Networks

In the era of data-driven applications, time series data is prolific, particularly in domains involving sensor-monitored systems. Such systems often produce a stream of multivariate time series data that can be leveraged to predict malfunctions and failures or, more broadly, anomalous events. Prediction of anomalies in time series data is a challenge that has been extensively pursued by researches, as failure forecast could save crucial resources in most applications. In this thesis, classical approaches to this problem will be discussed and a new approach that leverages graph convolutional neural networks will be presented. Such approach enables a much more in-depth analysis of multivariate time series since it allows to detect not only temporal correlations but also correlations between features. Moreover, cutting edge methodologies such as temporal convolutions and variational autoencoders will be exploited to build a powerful anomaly detection pipeline that is able to rival state-of-the art algorithms.