A Framework Enhanced by Mutual Information Cross Entropy for Time Series Anomaly Detection Under Noise

In this paper, an anomaly detection framework for time series is proposed, which is enhanced by mutual information cross entropy. Based on the prediction method, the mutual information cross entropy is used to select blocks that are most related to the prediction results. Considering the effects caused by various noise rates, a prediction module including a set of recurrent neural networks is trained and reserved under different noise environments, then a classification module consists of convolutional neural networks is used to choose suitable prediction model. Based on the errors between the predicted series and the test series, anomaly detection is implemented by Neyman Pearson criterion.