M.S. Candidate: Göksu Uzuntürk
Program: Data Informatics
Date: 25.06.2025 / 09:00
Place: A-212

Abstract: Range-based anomaly detection in multivariate time series plays a pivotal role in domains such as healthcare, industrial monitoring, finance, and cloud systems, where temporally extended faults often provide more actionable insights than isolated point anomalies. This study presents a transformer-based framework specifically designed to address the challenges of highly imbalanced multiclass range-based anomaly detection. To enhance temporal and semantic consistency, two post-inference strategies are incorporated: majority voting over overlapping multi-step predictions and a domain-informed transition masking mechanism that enforces realistic class transitions. These strategies contribute to output stability and more reliable diagnostics in scenarios governed by known operational constraints. The proposed method is evaluated on the Exathlon benchmark, demonstrating a notable improvement with a 24\% increase in F1 score of the weighted, binary anomaly detection evaluation framework.