Institutional Repository of Vilnius University
Anomalijų aptikimas daiktų interneto įrenginių duomenyse /
Abstract
dc:descriptionIn Internet of Things systems, sensor data is often used to monitor system state and support decision-making, which makes data reliability an important aspect. Inaccurate, atypical, or inconsistent sensor data may indicate device failure, connection failure, or another abnormal system behaviour. For this reason, anomaly detection in Internet of Things device data becomes an important part of data analysis, enabling timely identification of deviations from normal system behavior. This thesis analyzed the characteristics of Iintenet of Things data, types of anomalies, and the operating principles of anomaly detection methods. A scenario for a simulated smart irrigation system was defined for the study, where the system state is described by temperature, air humidity, soil moisture, water flow, and watering state data. Based on this scenario, a data generation and collection solution was implemented, enabling the collection of sensor data representing normal system behavior and the creation of an evaluation dataset injected with anomalies. The generated dataset simulates various types of deviations related to temperature spikes, inconsistencies between water flow and watering state, and changes in soil moisture over time. Three anomaly detection methods were applied in the study: K-nearest neighbors, local outlier factor, and isolation forest. The methods were applied using the initial sensor features and additional derived soil moisture features describing changes in values over time. The results of the methods were evaluated based on precision, recall, and the F1-score, as well as by analyzing how each method detects different types of anomalies. The results showed that derived soil moisture features improved the performance of all three methods. The best overall result was achieved by the local outlier factor method, with an F1-score of 0.640. This method detected 143 out of 150 injected anomalies and performed best in identifying collective anomalies related to a stuck soil moisture sensor. The K-nearest neighbors method was also suitable for detecting many anomalies, but its overall result was weaker. The isolation forest method performed the weakest in this study scenario. The obtained results showed that the quality of anomaly detection depends on the chosen method, the feature set used, and the nature of the anomalies.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Vaiginytė, Dovilė,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308338318&prefLang=en_US