Anomaly Detection in IoT Data

This work describes the performance-evaluation of various unsupervised classical machine learning algorithms in time series outlier detection. The aim is to test the robustness of known classical models that act as baselines in anomaly detection. IoT offers flexibility for various anomalies detection algorithms to be tested since the data collected is voluminous and the types of anomalies found are diverse. By deploying fine-tuned, long-established models, researchers can improve on the quality of the data they release from or use in various studies. This work also provides an insight into how time series data properties such as non-stationarity can affect anomaly detection and how operations such as windowing can be used to mitigate the effects and achieve desirable results. The experiments done show that, with some fine-tuning and data pre-processing, classical outlier detection methods' performance can be enhanced and utilized in IoT data quality control.