Deep Anomaly Detection: an experimental comparison of deep learning algorithms for anomaly detection in time series data

In the last decades, thanks to the advent of IoT (Internet of Things) and in general to the development of information technology, data have been doubling every year. Most of these data are collected in the form of Multivariate Time Series. The huge amount of time related information collected, allow us to monitor systems and identify deviations from the normal behavior, this is the well-known task of Anomaly Detection. Anomaly Detection is one of the most difficult tasks in Data Science due to issues and costs that needs to be afforded during the data collection and labelling processes. Anomaly Detection has been studied since the start of data science and statistics, but in the last years Deep learning advent opened the field of Deep Anomaly Detection and enabled the proliferation of more and more Nueral Network based algorithms which performances strictly depend on the nature of data. In this landscape it is quite difficult to choose a Deep Anomaly Detection algorithm to analyze some data we might have at disposal. While many surveys have been published about this argument, only a small number experimentally analyze the different methods reviewed, and most of the time they focus only on one kind of neural network architectures. In this work we provide a structured and experimental comparison of the state of the art Deep Anomaly Detection methods, in particular we divided algorithms proposed in literature based on the architecture type and then explored the best performing method for each architecture. The Chapter 3 reports a detailed description of the chosen methods which are: DeepAnt based on CNN, TanoGAN which make use of GANs, USAD an Auto-Encoder based model and Anomaly Transformer built with Transformer Netowrks. Those methods have been re-implemented by the author using python and the PyTorch framework library so that, the major practical output of this work is a python tool already provided with six datasets and those four algorithms which can be used to experimentally compare the performances of Deep Anomaly Detection methods. The tool is built in a modularized way so that it can be easily extended with other methods and datasets. As a final step the author performed a comparison of the methods presented providing the experimental results, a proper study of the metrics used in the comparison and conclusions about the identification of the best Deep Anomaly Detection model.