LLM-Assisted Policy-Adaptive Monitoring for Reliable IoT Sensor Data

This paper proposes an LLM-assisted self-adaptive monitoring system that enhances the reliability of environmental sensor data collected through a oneM2M-compliant IoT platform. The system integrates statistical anomaly detection, LLM-based interpretation, and rule-based policy adaptation within a continuous feedback loop, updating monitoring parameters―such as thresholds, window sizes, and check intervals―based on LLM-generated summaries rather than directly modifying sensor values. Experiments on a synthetic multi-sensor dataset demonstrate an absolute F1-score improvement of 9.0 percentage points over a non-adaptive LLM baseline, while a real-world case study using weather-station data confirms the system’s ability to detect missing-interval anomalies caused by pipeline disruptions and to adjust monitoring intervals accordingly. These results indicate that combining semantic reasoning with adaptive policy control can improve both the accuracy and interpretability of IoT data quality monitoring, offering a practical direction for autonomous and robust sensor data management.