Anomaly Detection in Streaming Time Series Based on Bounding Boxes

Anomaly detection in time series has been studied extensively by the scientific community utilizing a wide range of applications. One specific technique that obtains very good results is “HOT SAX”, because it only requires a parameter the length of the subsequence, and it does not need a training model for detecting anomalies. However, its disadvantage is that it requires the use of a normalized Euclidean distance, which in turn requires setting a parameter

ε

to avoid detecting meaningless patterns (noise in the signal). Setting an appropriate

ε

requires an analysis of the domain of the values from the time series, which implies normalizing all subsequences before performing the detection. We propose an approach for anomaly detection based on bounding boxes, which does not require normalizing the subsequences, thus it does not need to set

ε

. Thereby, the proposed technique can be used directly for online detection, without any a priori knowledge and using the non-normalized Euclidean distance. Moreover, we show that our algorithm computes less CPU runtime in finding the anomaly than HOT SAX in normalized scenarios.