Manuel Méndez, Alfredo Ibias and Manuel Núñez

Uncertainty is an ever present challenge in data analysis. In particular, it is important to detect, as precisely as possible, unforeseen phenomena. In this paper we study the usefulness of two deep learning based methods (CNN auto-encoder and BiLSTM auto-encoder) to detect anomalies in situations that can be defined in terms of time series. In order to evaluate our approaches, we consider traffic flow data and perform experiments in two orthogonal scenarios: a guided scenario (training only with data considered as ‘normal’ after a naïve labelling) and a basic scenario. Our results show that if we train the models using only the considered ‘normal’ data, the obtained models do not achieve good results because none of them are able to detect all type of abnormal data correctly. In contrast, both models can detect all type of time series anomalies when we consider the basic scenario.