Інтерпретовані реконструктивні методи виявлення аномалій у потокових даних інформаційних систем

Abstract

Modern information systems process large volumes of continuously arriving streaming data and require timely management decisions under strict time constraints. Real-time monitoring of such systems relies on automated anomaly detection methods, among which reconstructive approaches based on autoencoder encoder–decoder architectures are widely used to learn normal data patterns and detect anomalies through increased reconstruction error. However, these approaches face several limitations, including reconstruction inversion (the ability to reconstruct anomalous data), the inability to distinguish concept drift from true anomalies, and the lack of explainability for detected deviations. The proposed approach aims to reduce false alarms and improve detection accuracy while supporting operator situational awareness in monitoring systems such as telecommunications networks, cybersecurity monitoring platforms, and industrial control systems.