Detecção de anomalias em séries temporais variáveis com LSTM-RNNS

Abstract

The amount and availability of data and information grow continuously, and in parallel, the number of anomalies, corrections, intrusions, or cybernetics increases. This work proposes a combined model, which explores hybrid machine learning, to analyze data samples and detect anomalies and threats through scenario prediction. We do not always find large sets of real data - datasets - with balanced data samples, that is, many anomaly signals and normal states of a system, or environment, in equal proportions. The method was developed by exploring Long Short-Term Memory to work satisfactorily with different configurations and types of data set. It uses data grouping, classification, and fo- recasting techniques, analyzing states over time, defining normal behavior, and detecting anomalies. This approach also uses techniques to make it less necessary to have labeled and flagged data or any other indication of its use or sensitive information that cannot be revealed. The model allowed analyzing the anomalies, observing their advantages and limitations, with satisfactory models to detect them.