Scholarlink

Purpose <p>Accurate and computationally efficient prediction of bearing fault severity is essential for predictive maintenance and the reliability of rotating machinery. This study aims to present a lightweight diagnostic framework that enhances the extraction and discrimination of transient fault signatures.</p> Methods <p>The framework integrates the Smoothed Pseudo Wigner–Ville Distribution (SPWVD) for high-resolution time–frequency preprocessing with a lightweight Convolutional Neural Network (CNN, ~0.42M parameters, 0.1 GFLOPs). Comparative experiments evaluate SPWVD against the Short-Time Fourier Transform (STFT) and the Continuous Wavelet Transform (CWT) for extracting transient fault signatures.</p> Results <p>SPWVD outperforms STFT and CWT, delivering higher resolution and improved cross-term suppression. The proposed SPWVD-CNN model achieves 99.3% accuracy for 14-class fault severity on the Case Western Reserve University (CWRU) dataset and generalizes robustly with 94.4% accuracy on the Intelligent Maintenance System (IMS) dataset. The confusion matrix and per-class precision analyses confirm high reliability across fault types and severity levels.</p> Conclusion <p>Benchmarking against conventional classifiers and deep learning models demonstrates favourable accuracy, indicating that the SPWVD-CNN framework offers an efficient and reliable solution for bearing fault severity prediction.</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 Lightweight CNN Using Smoothed Pseudo-Wigner–Ville Distribution for Bearing Fault Severity Classification

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Vaibhav Shivhare,
  • Rajesh Kumar

摘要

Purpose

Accurate and computationally efficient prediction of bearing fault severity is essential for predictive maintenance and the reliability of rotating machinery. This study aims to present a lightweight diagnostic framework that enhances the extraction and discrimination of transient fault signatures.

Methods

The framework integrates the Smoothed Pseudo Wigner–Ville Distribution (SPWVD) for high-resolution time–frequency preprocessing with a lightweight Convolutional Neural Network (CNN, ~0.42M parameters, 0.1 GFLOPs). Comparative experiments evaluate SPWVD against the Short-Time Fourier Transform (STFT) and the Continuous Wavelet Transform (CWT) for extracting transient fault signatures.

Results

SPWVD outperforms STFT and CWT, delivering higher resolution and improved cross-term suppression. The proposed SPWVD-CNN model achieves 99.3% accuracy for 14-class fault severity on the Case Western Reserve University (CWRU) dataset and generalizes robustly with 94.4% accuracy on the Intelligent Maintenance System (IMS) dataset. The confusion matrix and per-class precision analyses confirm high reliability across fault types and severity levels.

Conclusion

Benchmarking against conventional classifiers and deep learning models demonstrates favourable accuracy, indicating that the SPWVD-CNN framework offers an efficient and reliable solution for bearing fault severity prediction.