Anomaly detection on time series using deep neural networks

Abstract

Despite significant advances in artificial intelligence and high-speed hardware, computers still do not have the same human capabilities for many cognitive activities. In the human brain, a large part of these activities is performed by the brain neocortex, one of which is detecting anomalies. In this study, we intend to use hierarchical temporal memory to recognize abnormal states in time series. Hierarchical temporal memory is a deep learning method that seeks to mimic the structural and algorithmic features of the brain neocortex. First, we implement this deep learning method, and then we present a solution to improve anomaly detection results. In our solution, in addition to the abnormality score that the network generates, we use the distance between the forecasted values and the actual values. We also change the classifier of this deep learning method to a support vector machine. The model presented achieved an F1 score of 0.350, which, compared to other deep learning methods, was one of the highest scores and the processing speed of this method was ten times faster than other methods.