Proposta de um algoritmo genérico de detecção de novidades em séries temporais utilizando modelos de previsão

Abstract

Novelties or anomalies on time series can be seen as unexpected values or a sequence of unexpected values when compared to a dataset considered to be normal. A time series novelty detection algorithm must be capable to classify each point of the time series as normal or novelty. There are many applications to the problem of time series novelty detection including fault detection on dynamic systems, fraud detection on financial systems and physiologic signals monitoring. This work proposes a generic novelty Detection algorithm based on a forecasting model. The algorithm is generic because it does not define the forecasting model to be used. Two algorithm instantiations are proposed, the first one, based on a statistical model and the second one based on a neural model. The algorithm is used to screen obstructive sleep apnea through electrocardiogram monitoring and to detect faults on a dynamic system monitoring some of the system variables. The results achieved on both problems are near to the results found on literature.