Anomaly Detection in Internet of Things (IoT) Time Series Data: A Comparative Study of Various Techniques

Abstract

Anomaly detection plays an integral role in a broad range of applications within the Internet of Things (IoT), such as preventive maintenance, health monitoring, fraud detection, and fault prediction. This study undertakes a comprehensive exploration of the methods commonly used for anomaly detection in IoT time series data. These methods encompass Statistical Techniques, Isolation Forest, Autoencoder Neural Networks, and Long Short-Term Memory Units (LSTMs), each with their unique strengths and challenges. Statistical techniques, such as ARIMA, ETS, and STL, model the regular pattern of a time series via a stochastic model, highlighting anomalies as instances that deviate from this model. The Isolation Forest algorithm, on the other hand, isolates anomalies based on their shorter average path lengths in an ensemble of Isolation Trees. Autoencoders and LSTMs, as types of artificial neural networks, detect anomalies via high reconstruction error and significant deviation from predicted values, respectively. The research also acknowledges the applicability of other methods such as K-means clustering, DBSCAN, and XGBoost according to the specific requirements of IoT data. Selection of an appropriate model depends largely on the data characteristics and the particular use case, with data properties including multivariate or univariate nature, presence of trends or seasonality, and type of anomalies playing a crucial role.