EE Seminar: Unsupervised Detection and Interpretation of Anomalies in Commercial Microwave Links
https://tau-ac-il.zoom.us/j/82319707943?pwd=x4ZrJdj5OuBOu668jPQVFxBvWHB8OR.1
Electrical Engineering Systems ZOOM Seminar
Speaker: Adi Green
M.Sc. student under the supervision of Prof. Hagit Messer-Yaron
Wednesday, 11th February 2026, at 15:00
Unsupervised Detection and Interpretation of Anomalies in Commercial Microwave Links
Abstract
Commercial Microwave Links (CMLs) serve as the backbone of cellular communication networks and have emerged as opportunistic environmental sensors capable of detecting rainfall through signal attenuation. However, CML-based monitoring faces fundamental challenges: operational data is inherently contaminated with unknown mixtures of environmental effects and technical disturbances, and while supervised methods have shown success for rainfall detection, ground truth labels for other phenomena such as technical faults and hardware degradation remain largely undocumented. This thesis presents a fully unsupervised machine learning framework for detecting and characterizing anomalies in CML signals without labeled data. The methodology integrates three components: (1) an autoencoder-based anomaly detector utilizing hysteresis thresholding to preserve temporal consistency; (2) a feature engineering module extracting 74 features across spatial, signal morphology, and temporal domains, refined via Random Forest feature selection; and (3) a K-means clustering mechanism optimized by silhouette scores. The framework was validated on 647 CMLs across four tropical Caribbean regions over 1.5 years. The optimized model achieved binary silhouette scores of 0.38–0.39 for anomaly detection and 0.27–0.28 for multi-class clustering, demonstrating reasonable generalization despite training on contaminated operational data. To explore physical interpretability, results were cross-referenced with external ground truth from the NOAA Integrated Surface Database. The analysis revealed potentially meaningful functional separation: one cluster exhibited near-zero rain gauge association and a "single-fault-per-link" pattern, suggesting isolation of hardware faults, while other clusters showed elevated rain associations (3.2% to 5.3%) and high recurrence rates, consistent with environmental phenomena. These interpretations represent working hypotheses requiring further validation. The key contribution is a complete, fully unsupervised framework for CML anomaly detection that operates without labeled data dependencies, with modular design enabling future improvements. This framework opens new possibilities for autonomous CML network monitoring in regions lacking meteorological infrastructure.
-סמינר זה ייחשב כסמינר שמיעה לתלמידי תואר שני ושלישי-
This Seminar Is Considered A Hearing Seminar For Msc/Phd Students-
כדי לקבל קרדיט שמיעה יש לחתום שם מלא ומספר ת.ז. בצ'ט
