Scholarlink

<p>Uncrewed aerial vehicles (UAV(s)) are emerging as agile non-terrestrial network (NTN) nodes that extend 5G/6G coverage to remote farms, disaster zones, and shipping lanes. However, on-board ML models carried by UAV(s) degrade under concept drift, leading to missed events and false alarms. We propose a real-time drift-alarm framework (<span>RTDD</span>) that pairs a deployed classifier with a lightweight clustering model. On board, <span>RTDD</span> cross-checks classifier outputs against cluster membership to infer error on unlabelled streams and signals drift when the discrepancy between the on-board classifier-accuracy estimate (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({\textbf {CA}}\)</EquationSource> </InlineEquation>) and the classifier–cluster agreement rate (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\({\textbf {ClA}}\)</EquationSource> </InlineEquation>) exceeds a conservative discrepancy threshold. This thresholding preserves sensitivity to both abrupt and gradual shifts while suppressing spurious alarms. Empirically, agreement between classifier outputs and cluster membership provides a robust, label-free drift signal. Using this signal, our method achieves similar detection timeliness to state-of-the-art label-based approaches while reducing the false-alarm rate under true-label delay.</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

A Drift-alarm Framework for NTN–UAV Nodes: Robust, Self-healing ML Models via Classifier–cluster Consistency

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Salman Ahmed,
  • Nayyer Masood,
  • Masood Ur-Rehman

摘要

Uncrewed aerial vehicles (UAV(s)) are emerging as agile non-terrestrial network (NTN) nodes that extend 5G/6G coverage to remote farms, disaster zones, and shipping lanes. However, on-board ML models carried by UAV(s) degrade under concept drift, leading to missed events and false alarms. We propose a real-time drift-alarm framework (RTDD) that pairs a deployed classifier with a lightweight clustering model. On board, RTDD cross-checks classifier outputs against cluster membership to infer error on unlabelled streams and signals drift when the discrepancy between the on-board classifier-accuracy estimate ( \({\textbf {CA}}\) ) and the classifier–cluster agreement rate ( \({\textbf {ClA}}\) ) exceeds a conservative discrepancy threshold. This thresholding preserves sensitivity to both abrupt and gradual shifts while suppressing spurious alarms. Empirically, agreement between classifier outputs and cluster membership provides a robust, label-free drift signal. Using this signal, our method achieves similar detection timeliness to state-of-the-art label-based approaches while reducing the false-alarm rate under true-label delay.