Authors: Debanjana Nayak; Harry Perros
Addresses: Department of Computer Science, North Carolina State University, Engineering Building II, 890 Oval Dr., Raleigh, NC 27606, USA ' Department of Computer Science, North Carolina State University, Engineering Building II, 890 Oval Dr., Raleigh, NC 27606, USA
Abstract: Fast identification of faulty sensors is necessary for guaranteeing their robust functions in diverse applications ranging from extreme weather prediction to energy saving to healthcare. We present an automated machine-learning based framework that can detect anomalies of temperature sensor data in real-time. We adopted a purely temporal approach that utilises a univariate time-series (UTS) generated by a single sensor. The framework divides the UTS into subsequences, models each subsequence stochastically as an autoregressive function, and finally mines the function parameters with a one-class support vector machines (OC-SVM) that classifies any outlier as an anomaly. Extensive experimentation showed that the framework identifies both normal and anomalous data correctly with high degrees of accuracy.
Keywords: UTS; univariate time-series; anomaly detection; temperature sensors; OC-SVMs; one-class support vector machines; autoregression.
International Journal of Sensor Networks, 2020 Vol.34 No.3, pp.137 - 152
Received: 22 Apr 2020
Accepted: 24 Apr 2020
Published online: 26 Oct 2020 *