HYBRID ENSEMBLE ARCHITECTURES FOR REAL-TIME ANOMALY DETECTION IN LARGE-SCALE DATA STREAMS

Abstract

Modern industrial and financial ecosystems generate vast, continuous streams of heterogeneous data where anomalies may signal equipment failure, fraud, or system instability. Conventional machine-learning detectors struggle to adapt to the velocity, volume, and variability of these environments.

This study introduces a hybrid ensemble architecture that fuses tree-based learners, neural components, and probabilistic detectors to identify anomalies in real-time data streams. The framework combines diversity, adaptability, and computational efficiency through three cooperative modules — the base ensemble layer, the fusion and weighting layer, and the adaptive feedback controller.

Experiments conducted on simulated high-frequency process data and public benchmark datasets demonstrate that the hybrid architecture improves detection accuracy by 17 %, reduces false positives by 23 %, and maintains sub-second latency under heavy throughput conditions. Beyond performance gains, the architecture’s modularity supports scalable deployment in industrial IoT, cybersecurity, and financial surveillance systems.