Scholarlink

Abstract <p>Fast and reliable anomaly detection in rural and urban macro wireless channels is impeded by diverse multipath and mobility patterns. A scalable, low-latency pipeline is presented to preserve critical variance while avoiding multicollinearity preprocessing. An Elastic Net-guided selection is followed by a kernel-optimized dimensionality reduction that removes the variance inflation factor (VIF) layer. A hybrid detector is then employed, combining Isolation Forest for global outliers with density-based spatial clustering of applications with noise (DBSCAN) for local validation. To avoid dependence on grid search or cross-validation, a kernel-width sampling strategy is adopted at the kernel principal component analysis (KPCA) stage. Experiments are conducted under 3GPP TR 38.901 configurations with SAGE/QuaDRiGa-derived features. Seven features are reduced to four and projected to two components, yielding a total explained variance of 76.5%, which exceeds prior two-component baselines (71–76%). An anomaly rate of 5.0% is observed across rural macro-cell (RMa) and urban macrocell (UMa) scenarios under line-of-sight and non-line-of-sight conditions. Accuracy of 97–100% is achieved across six classifiers with reduced pipeline delay. To the authors’ knowledge, the accompanying package constitutes one of the first open-source, end-to-end implementations for wireless anomaly detection that integrates VIF-free KPCA with grid-search–free kernel-width sampling and hybrid global–local validation. The package (<i>wireless-anom</i>) has been archived on Zenodo (DOI reserved: <a href="https://zenodo.org/records/17096385?token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6IjMyNzI3ODZlLWMyMWItNGRkMS05ZWVlLWNhZTA0MDI4MWQ0OSIsImRhdGEiOnt9LCJyYW5kb20iOiJiNTJiODlkYzVhYmZhZWNmNjgyNmE1ZjZhY2Q1M2ZmMyJ9.m-X5-fDe2G_IDXP3m1xqymnAaTT3Q-NP9PLoZqTVsq0P2kLmk-mcaqoVsiNLeKfIqip1kswkIz5qpeo87XGhpA">10.5281/zenodo.17096385</a>) with containers and scripts.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AI-driven anomaly detection for wireless networks: a scalable and efficient approach using optimized kernel PCA and isolation forest

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Ahmed Métwalli,
  • Waleed K. Badawi

摘要

Abstract

Fast and reliable anomaly detection in rural and urban macro wireless channels is impeded by diverse multipath and mobility patterns. A scalable, low-latency pipeline is presented to preserve critical variance while avoiding multicollinearity preprocessing. An Elastic Net-guided selection is followed by a kernel-optimized dimensionality reduction that removes the variance inflation factor (VIF) layer. A hybrid detector is then employed, combining Isolation Forest for global outliers with density-based spatial clustering of applications with noise (DBSCAN) for local validation. To avoid dependence on grid search or cross-validation, a kernel-width sampling strategy is adopted at the kernel principal component analysis (KPCA) stage. Experiments are conducted under 3GPP TR 38.901 configurations with SAGE/QuaDRiGa-derived features. Seven features are reduced to four and projected to two components, yielding a total explained variance of 76.5%, which exceeds prior two-component baselines (71–76%). An anomaly rate of 5.0% is observed across rural macro-cell (RMa) and urban macrocell (UMa) scenarios under line-of-sight and non-line-of-sight conditions. Accuracy of 97–100% is achieved across six classifiers with reduced pipeline delay. To the authors’ knowledge, the accompanying package constitutes one of the first open-source, end-to-end implementations for wireless anomaly detection that integrates VIF-free KPCA with grid-search–free kernel-width sampling and hybrid global–local validation. The package (wireless-anom) has been archived on Zenodo (DOI reserved: 10.5281/zenodo.17096385) with containers and scripts.

Graphical abstract