Machine Learning with stream processing engines for IoT applications
Mariona Carós Roca · UPCommons institutional repository (Universitat Politècnica de Catalunya) · 2017
The Internet of Things (IoT) enables to connect multiple devices for providing a certain service, consequently huge amount of data is generated in time, known as time series. This phenomenon presents unique challenges in defining the data behavior and detecting anomalies. In this thesis, we present an appropriate method for defining the normal behavior of the time series and detection of anomalies. We generate a daily periodic data set of time series based on the analysis of an energy consumption real data. Then, by observing the input data, assumed to be independent from an unknown probability distribution, we define the normal behavior. The description of the data distribution is obtained by certain statistics and a Marked Point Process of change points. We develop techniques for detecting the anomalies and providing the type of anomaly as well, using a Multiple Hypothesis Testing . Finally, we present some experiments with the synthetic and real time series.