Scholarlink

<p>This paper presents an innovative engineering informatics framework for identifying subtle, low-magnitude anomalies in complex time series, providing a critical tool for the proactive safety monitoring of infrastructure systems. The proposed technique demonstrates superior resilience compared to existing literature, specifically on data where severe noise coexists with non-linear dynamics. The foundation involves a tailored Support Vector Regressor for refined filtering that actively suppresses intrinsic non-polynomial trends. Following trend-removal, anomaly detection, serving as a condition-based maintenance trigger, is executed via a bifurcated strategy. This integrates an analysis of inter-tree variance in a customized Random Forest model sequence with a Root Mean Squared Error assessment. By quantifying the breakdown in model consensus, the inter-tree variance acts as an uncertainty-aware metric that enhances diagnostic reliability. This synergistic approach distinguishes minute deviations to mitigate risks of catastrophic failure. Applying this methodology to Water Distribution Network steady-pressure data demonstrates this capability, enabling the detection of incipient leaks at their earliest stages. Stringent validation on a realistic synthetic dataset confirms the methodology successfully identifies small-scale leaks (≈11&#xa0;l/s), representing less than 0.2% of the network's total baseline flow, previously undetectable by state-of-the-art methods like Convolutional Neural Networks and Long Short-Term Memory neural networks. This research contributes to reliability engineering by providing a computationally efficient framework for identifying nascent failure modes in critical smart infrastructures.</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

An uncertainty-aware reliability framework for incipient failure detection in critical infrastructure: application to water distribution networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月2日

查看全文
  • Elvio Damonti,
  • Giancarlo Bernasconi

摘要

This paper presents an innovative engineering informatics framework for identifying subtle, low-magnitude anomalies in complex time series, providing a critical tool for the proactive safety monitoring of infrastructure systems. The proposed technique demonstrates superior resilience compared to existing literature, specifically on data where severe noise coexists with non-linear dynamics. The foundation involves a tailored Support Vector Regressor for refined filtering that actively suppresses intrinsic non-polynomial trends. Following trend-removal, anomaly detection, serving as a condition-based maintenance trigger, is executed via a bifurcated strategy. This integrates an analysis of inter-tree variance in a customized Random Forest model sequence with a Root Mean Squared Error assessment. By quantifying the breakdown in model consensus, the inter-tree variance acts as an uncertainty-aware metric that enhances diagnostic reliability. This synergistic approach distinguishes minute deviations to mitigate risks of catastrophic failure. Applying this methodology to Water Distribution Network steady-pressure data demonstrates this capability, enabling the detection of incipient leaks at their earliest stages. Stringent validation on a realistic synthetic dataset confirms the methodology successfully identifies small-scale leaks (≈11 l/s), representing less than 0.2% of the network's total baseline flow, previously undetectable by state-of-the-art methods like Convolutional Neural Networks and Long Short-Term Memory neural networks. This research contributes to reliability engineering by providing a computationally efficient framework for identifying nascent failure modes in critical smart infrastructures.