An Interpretable Agent-Assisted Pipeline for Statistical Anomaly Detection in IoT Temperature Time Series

The research presents an interpretable framework which detects anomalies in IoT temperature time-series data with low complexity for use in edge environments that lack resources. The proposed solution uses three traditional statistical filters which include Hampel and Interquartile Range (IQR) and Z-Score to build an agent-assisted decision layer which selects the best method through a multi-criteria cost function. The framework runs tests on a structured synthetic dataset which contains seven different anomaly tests and on an actual IoT dataset which was gathered from eight separate sensor points. The researchers use standard anomaly detection metrics which include precision and recall and F1-score and false positive rate to conduct their complete evaluation. The proposed method is tested against two machine learning baseline methods which are Isolation Forest and One-Class Support Vector Machine (OC-SVM). The results show that the agent-assisted method achieves detection results which match industry standards while showing high interpretability and low processing needs. The framework demonstrates its ability to function in actual IoT environments through its use of authentic real-world data, and also basic statistical techniques together with an adjustable decision system create a strong and understandable method to detect anomalies in IoT sensing systems.