Soluções em detecção de anomalias em séries temporais multivariadas utilizando modelos preditivos

Abstract

Time-series anomalies detection is a fast-growing area of study, due to the exponential growth of new data produced by sensors in many different contexts as the Internet of Things (IoT). Many predictive models have been proposed, and they provide promising results in differentiating normal and anomalous points in a time-series. In this work, we provide three contributions. We aim to identify and combine the best models for detecting anomalies in time-series, so that the different strategies or parameters of the models can contribute to the time series analysis by proposing a model-centered ensemble called TSPME-AD (stands for Time Series Prediction Model Ensemble for Anomaly Detection). TSPME-AD uses state-of-the-art predictive models, combining their anomaly scores using a weighted function. Other contributions are a dynamic window breaking technique based on scanning thresholds, leveraging the periodicity and shape of the series to generate windows that aid in training and testing models, and a new auto-encoder predictive model. The effectiveness of our proposals is analyzed using two real-world time-series datasets, a year of power demand data, and the MIT electrocardiogram database. We show that our ensemble technique improves on the F1 score up to 22% on the best score of the individual models composing the ensemble, with our specific combination function improving on simpler functions on up to 13% F1 score increase. We also show that our new auto-encoder architecture, combined with the new window breaking technique, can have an up to 25% F1 score increase compared to another proposed auto-encoder technique, and a 64% score increase over a stacked LSTM model.