Edge AI-based Time Series Anomaly Detection addresses a learning framework for reliably detecting anomalies in time-series data within edge environments with limited computational resources. This study goes beyond simply improving accuracy, aiming to simultaneously ensure decision stability and interpretability while accounting for data quality variations and uncertainty. In particular, it designs a practically deployable anomaly detection system by combining lightweight model architectures with uncertainty-aware decision making.

• Problem Anomaly detection degradation under edge constraints and data instability
• Approach Fusion-based decision framework combining anomaly scores and uncertainty
• Model Lightweight Teacher–Student-based Deep Anomaly Fusion model
• Analysis Performance–computation trade-off and decision stability analysis
• Goal Real-time, lightweight, and explainable anomaly detection system