An efficiency-driven anomaly detection pipeline for mobile networks
摘要
In recent years, the growth of new wireless technologies and the emergence of new requirements in applications and services are progressively changing the way networks are used. Network usage analysis has become a useful tool that enables network administrators to identify anomalies, providing them with the ability to define infrastructure requirements and avoid network saturation during peak hours. We present a modular pipeline for anomaly detection that models normal network behaviour from stable baseline periods. New data are compared to this model using an orthogonal-subspace projection method, producing clear and interpretable anomaly scores. Scores are normalised to a dimensionless severity measure, enabling focused detection of extreme events. This approach supports scalable, parallel processing and ensures consistent, reproducible results across different hardware platforms. We tested the same code and configuration on both a compact ARM-based system on chip with unified memory and a traditional high-performance computing node. Across three Apple Silicon generations, time-to-solution improves by up to