Rilevamento di Anomalie in Serie Temporali usando le Shapelets

In time series classification literature, shapelets are subsequences able to discriminate the different classes. It has been shown that classifiers achieve high accuracy when taking in input the distances from the time series and a properly chosen set of shapelets. Additionally, these subsequences can be easily compared visually to input signals: this characteristic makes such algorithms naturally interpretable. More recent methods propose to parametrize the shapelets instead of searching them through all the possible subsequences, which is computationally expensive. The optimal shapelets are learned through the minimisation of a loss function. This approach typically sacrifices their interpretability in favour of a higher accuracy. This thesis has been carried on in collaboration with Addfor Industriale S.r.l. towards the goal of developing shapelets-based algorithms for unsupervised anomaly detection in univariate and multivariate time series datasets. We propose and analyse three algorithms and compare their results on five benchmark datasets, focusing on the difference between searching and learning approaches. The experiments confirm the main advantages of shapelets methods, which work well when recurring short patterns distinguish the shape of normal and anomalous time series. In particular, the proposed shapelets-learning algorithm manages to learn different discriminating shapelets, while maintaining their interpretable characteristic.