Scholarlink

<p>This study introduces an automatic gross error identification method for hydraulic monitoring data based on a multi-level L-Transformer framework. The proposed architecture integrates Liquid Neural Networks for short-term disturbance detection with Transformer modules for long-term dependency modeling, while adaptive threshold learning and integrated decision logic establish a closed-loop self-identification mechanism. The method was validated using over 15 million records from 635 sensors deployed along a large water diversion tunnel in central China. A strict cross-point partition was adopted, with 631 points for training and four independent points for testing, thereby eliminating any risk of data leakage. Even with only 30% of training data, the model achieved 99.26% accuracy on unseen targets, attributed to the clear distributional differences between gross errors and normal data. Robustness was further demonstrated under label-noise perturbations, confirming that the results are not attributable to overfitting. With the full dataset, the model achieved 99.87% accuracy, significantly outperforming established baselines including PLSR (3-sigma), LOF, Random Forest, LNN, Transformer, and ensemble methods. In addition, the framework demonstrated real-time feasibility (~ 0.5 ms per sample), underscoring its strong potential for online monitoring and edge deployment in hydraulic systems.</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

Based on Multi-level Analysis: A Self-Identification Model for Gross Errors in Hydraulic Monitoring Data

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • Yilun Wei,
  • Yi Liu,
  • Xinxin Jin,
  • Mengyao Shen,
  • Hui Li,
  • Ding Nie

摘要

This study introduces an automatic gross error identification method for hydraulic monitoring data based on a multi-level L-Transformer framework. The proposed architecture integrates Liquid Neural Networks for short-term disturbance detection with Transformer modules for long-term dependency modeling, while adaptive threshold learning and integrated decision logic establish a closed-loop self-identification mechanism. The method was validated using over 15 million records from 635 sensors deployed along a large water diversion tunnel in central China. A strict cross-point partition was adopted, with 631 points for training and four independent points for testing, thereby eliminating any risk of data leakage. Even with only 30% of training data, the model achieved 99.26% accuracy on unseen targets, attributed to the clear distributional differences between gross errors and normal data. Robustness was further demonstrated under label-noise perturbations, confirming that the results are not attributable to overfitting. With the full dataset, the model achieved 99.87% accuracy, significantly outperforming established baselines including PLSR (3-sigma), LOF, Random Forest, LNN, Transformer, and ensemble methods. In addition, the framework demonstrated real-time feasibility (~ 0.5 ms per sample), underscoring its strong potential for online monitoring and edge deployment in hydraulic systems.