Anomaly detection in time series

Anomaly detection (AD) is the discovery of the observations which does not conform with the rest of the observations. The types of anomalies and their occurrences that exist in the data set are tried to be determined. On the other hand, time series structures have dynamic structures, which are evolving over time, and in such structures, observations will be affected by previous observations. This thesis focuses on the anomaly detection process under time series structures. This problem is not always straightforward because the definition of anomaly could change with the context of the dynamic structure and anomaly detection process in the system could interfere with the intense noises at the observations. In this thesis, we try to identify anomalies in the sub-sequences of the streaming data. When doing so, we also want to discriminate the anomalies in the system with the faulty observations. Therefore we investigate collective anomalies in the data. We propose both statistical inference methods and deep learning approaches for such type of anomaly detection in time series (ADTS) problem. We use a Gaussian mixture model (GMM) and a customized hidden Markov model (HMM) as statistical approaches, while we use Recurrent Neural Networks (RNN) and Long Short-Term Memories (LSTM) as deep learning approaches. Except for GMMs, we take into account the sequential structures of data sets in the models proposed above. We apply our methodologies to the Borusan wind turbines data and we compare the model results with the experiments we performed on this dataset.