Scholarlink

<p>This study investigates bearing fault diagnosis using compact neural models and interpretable features. We compare time-signal statistics with a Wavelet Packet Transform (WPT) representation (level-3, Fejér–Korovkin), where each 1-s window is mapped to an 8-dimensional relative-energy vector. Time-signal analysis yields ~ 80–87% accuracy depending on signal type, with feature selection (MRMR) offering limited gains. In contrast, WPT on laboratory, noise-free vibration achieves &gt; 99% test accuracy using a small multilayer perceptron (MLP). To assess robustness under realistic interference, we generate noisy datasets with multiple mechanisms—white/pink/brown, Bernoulli–Laplace impulsive (BLIN), shot noise (SN) across − 10… + 10&#xa0;dB SNR, and repeating-pulse noise (RPN) spanning 1–25.6&#xa0;kHz. Under white gaussian noise (WGN)-only, accuracy ranges from ~ 70% (− 10&#xa0;dB) to ~ 93% (+ 10&#xa0;dB), averaging ~ 86% across all SNRs. To reduce the lab-to-field gap, we introduce a denoising autoencoder (DAE) trained in the DWPT space to map noisy features back to their clean templates. With denoising, cross-validated accuracy reaches ~ 93% (K-fold) and held-out test accuracy ~ 87% on mixed-noise signal, while preserving a small footprint (&lt; 1.3&#xa0;MB) and millisecond-level per-window latency suitable for edge deployment.</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

Artificial neural networks and time–frequency analysis for bearing fault detection under noisy conditions

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Bartu Işık,
  • Kubilay Kılıç,
  • Arda Tanikyan,
  • Ali Bahadır Olcay

摘要

This study investigates bearing fault diagnosis using compact neural models and interpretable features. We compare time-signal statistics with a Wavelet Packet Transform (WPT) representation (level-3, Fejér–Korovkin), where each 1-s window is mapped to an 8-dimensional relative-energy vector. Time-signal analysis yields ~ 80–87% accuracy depending on signal type, with feature selection (MRMR) offering limited gains. In contrast, WPT on laboratory, noise-free vibration achieves > 99% test accuracy using a small multilayer perceptron (MLP). To assess robustness under realistic interference, we generate noisy datasets with multiple mechanisms—white/pink/brown, Bernoulli–Laplace impulsive (BLIN), shot noise (SN) across − 10… + 10 dB SNR, and repeating-pulse noise (RPN) spanning 1–25.6 kHz. Under white gaussian noise (WGN)-only, accuracy ranges from ~ 70% (− 10 dB) to ~ 93% (+ 10 dB), averaging ~ 86% across all SNRs. To reduce the lab-to-field gap, we introduce a denoising autoencoder (DAE) trained in the DWPT space to map noisy features back to their clean templates. With denoising, cross-validated accuracy reaches ~ 93% (K-fold) and held-out test accuracy ~ 87% on mixed-noise signal, while preserving a small footprint (< 1.3 MB) and millisecond-level per-window latency suitable for edge deployment.