Authors:
Aafan Toor
;
Jia-Chun Lin
;
Ming-Chang Lee
and
Ernst Gran
Affiliation:
Department of Information Security and Communication Technology, Norwegian University of Science and Technology, Gjøvik, Norway
Keyword(s):
Internet of Things, Smart Home, Long Short-Term Memory (LSTM), Online Learning, Time Series, Anomaly Detection, Sliding Window.
Abstract:
In the context of time series data, a contextual anomaly is considered an event or action that causes a deviation in the data values from the norm. This deviation may appear normal if we do not consider the timestamp associated with it. Detecting contextual anomalies in real-world time series data poses a challenge because it often requires domain knowledge and an understanding of the surrounding context. In this paper, we propose UoCAD, an online contextual anomaly detection approach for multivariate time series data. UoCAD employs a sliding window method to (re)train a Bi-LSTM model in an online manner. UoCAD uses the model to predict the upcoming value for each variable/feature and calculates the model’s prediction error value for each feature. To adapt to minor pattern changes, UoCAD employs a double-check approach without immediately triggering an anomaly notification. Two criteria, individual and majority, are explored for anomaly detection. The individual criterion identifies
an anomaly if any feature is detected as anomalous, while the majority criterion triggers an anomaly when more than half of the features are identified as anomalous. We evaluate UoCAD using an air quality dataset containing a contextual anomaly. The results show UoCAD’s effectiveness in detecting the contextual anomaly across different sliding window sizes but with varying false positives and detection time consumption.
(More)