Scholarlink

<p>Fault detection in covered conductors is critical, yet field data often exhibit high background noise. Traditional feature engineering methods can help, but many rely on a single feature, limiting their ability to support reliable detection in noisy raw signals. We aim to develop a multi-scale, learning-based ensemble for fault detection in covered conductors, leveraging features at multiple scales to represent raw signals more faithfully and improve detection accuracy. The measurements are first preprocessed to identify the pulses and the surrounding waveforms. Then, statistical features are extracted through feature engineering, followed by data augmentation. Next, the proposed Patch Encoder is employed to extract the underlying information within waveforms, while a Lightweight Encoder is used to capture the well-designed statistical features. A Learnable Ensemble Mechanism is then used to combine the outputs of Patch Encoder and the Lightweight Encoder, achieving improved performance. Finally, experimental results on the VSB ENET dataset validate the effectiveness of MsLEM, which achieves a Matthews Correlation Coefficient (MCC) of 0.757, outperforming existing state-of-the-art methods. Under challenging conditions with a 15 dB additional SNR, the model maintains an MCC of 0.701, reflecting only a 7.4% performance degradation. Furthermore, on a Simulink-based three-phase dataset with power swings, load disturbances, 20 dB SNR additive white Gaussian noise, and random high-frequency damped oscillation events, MsLEM achieves an MCC of 0.703, confirming its robustness.</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

MsLEM: multi-scale learning-based ensemble model for fault detection of covered conductors

  • 期刊论文

  • 发表日期：

    2026年4月16日

查看全文
  • Liyang Xu,
  • Dezheng Wang

摘要

Fault detection in covered conductors is critical, yet field data often exhibit high background noise. Traditional feature engineering methods can help, but many rely on a single feature, limiting their ability to support reliable detection in noisy raw signals. We aim to develop a multi-scale, learning-based ensemble for fault detection in covered conductors, leveraging features at multiple scales to represent raw signals more faithfully and improve detection accuracy. The measurements are first preprocessed to identify the pulses and the surrounding waveforms. Then, statistical features are extracted through feature engineering, followed by data augmentation. Next, the proposed Patch Encoder is employed to extract the underlying information within waveforms, while a Lightweight Encoder is used to capture the well-designed statistical features. A Learnable Ensemble Mechanism is then used to combine the outputs of Patch Encoder and the Lightweight Encoder, achieving improved performance. Finally, experimental results on the VSB ENET dataset validate the effectiveness of MsLEM, which achieves a Matthews Correlation Coefficient (MCC) of 0.757, outperforming existing state-of-the-art methods. Under challenging conditions with a 15 dB additional SNR, the model maintains an MCC of 0.701, reflecting only a 7.4% performance degradation. Furthermore, on a Simulink-based three-phase dataset with power swings, load disturbances, 20 dB SNR additive white Gaussian noise, and random high-frequency damped oscillation events, MsLEM achieves an MCC of 0.703, confirming its robustness.