Deep learning models for anomaly detection in time series

Recent advances in technology allow us to collect a large amount of data over time in diverse fields and the amount of data transferred exceeds the human ability to study it manually. Generally, standard statistical approaches assume that the samples are generated by a specific statistical model and do not scale well with the amount of data and thus automated data analysis becomes necessary. On the other hand, machine learning methods consider the data generation process as a black box and try to learn from the input only. Moreover, in many application fields, such as economic, healthcare and security, there is the need to have fast computations but at the same time receive a reliable result. Traditional models use supervised machine learning algorithms but, in the context of applications, collecting and annotating such large-scale datasets is difficult, time-consuming or even too expensive, and it requires domain knowledge from experts in the field. Therefore, anomaly detection has been such a great challenge for researchers and practitioners. Anomaly detection is referred to as the process of detecting anomaly data instances. The definition of an anomaly depends on the task and domain but, most of the time, it is an instance that significantly deviates from the others. In this thesis, the focus is on deep learning models for anomaly detection in time series. In the first part, a general overview of the anomaly detection task is provided and the properties and the definition of the time series are presented. Then, in the second part, various state of the art anomaly detection algorithms are discussed. Moreover, I present two new approaches, along with a comparison with the classic methods. In the third part, experiments carried out with different datasets and different architectures are shown. Furthermore, I provide some improvements to the presented methods. In detail, the experiments are made with two public datasets and one on damage detection in industrial composite structures. These datasets have different properties in order to show how the discussed methods perform in different situations. In the end, the results show the ability of the proposed models to detect anomalous patterns in time series from different fields of application while providing structured and expressive data representations.