The primary purpose of the ARTEMYS (Artificial in TElligence and Machine learning validation sYStem) project, funded under the NextGenerationEU framework within the RETURN project (Spoke VS1), is to establish an automated, real-time validation service for hydro-meteorological sensor data. In environmental monitoring networks, raw sensor readings are frequently corrupted by transient or persistent disturbances across the data acquisition chain, necessitating an automated mechanism to assign confidence scores and filter out errors before data reaches research and warning systems. To achieve this, ARTEMYS utilizes a spatial clustering framework combined with an innovative dual-model machine learning strategy. Experimental evaluation revealed that standalone models—specifically Random Forest classifiers and deep learning Anomaly Transformers—exhibited high recall by capturing almost all real anomalies, but suffered from extremely low precision, generating thousands of false positives. ARTEMYS addresses this bottleneck through an ensemble consensus mechanism that flags a reading as anomalous only when both models agree, leveraging their distinct architectural strengths to eliminate false alarms while maintaining high detection sensitivity.
Following an initial evaluation of state-of-the-art anomaly detection techniques, the team tested the models on real-world rainfall datasets provided by ARPAL Liguria. In a test dataset containing over 1.4 million samples and 444 genuine anomalies, individual models generated up to 2,595 false positives. The ensemble approach dramatically improved performance, correctly identifying 442 out of the 444 true anomalies while reducing false positives to just 16. In the engineering and deployment phases, the research logic was refactored into a standalone Python pipeline executing spatial clustering, parallel dual-model inference, and ensemble filtering.
2024–2025