Scholarlink

<p>Accurate fault diagnosis of rolling bearings is essential for ensuring the reliable operation of rotating machinery, particularly under noisy environments and complex operating conditions. To address this challenge, this paper presents a robust and noise-resilient bearing fault diagnosis framework that integrates advanced signal processing with hybrid deep learning techniques. Initially, raw vibration signals are decomposed using ensemble empirical mode decomposition (EEMD) to extract intrinsic mode functions that effectively capture fault-related characteristics under nonstationary conditions. Correlation coefficient-based analysis is then employed to identify the most informative components, thereby suppressing irrelevant and noise-dominated features. The selected features are normalized and encoded into two-channel Gramian angular field (GAF) representations, preserving temporal correlations and fault dynamics in a two-dimensional image domain. A convolutional neural network (CNN) embedded with an improved residual shrinkage network (IRSN) is subsequently utilized to perform adaptive noise suppression through learnable soft-thresholding while extracting discriminative spatial features. To further enhance representation capability and capture long-range dependencies under varying fault conditions, a Transformer encoder is incorporated after the CNN–IRSN stages, enabling effective modeling of global contextual information. The learned deep features are finally classified using an extreme learning machine (ELM), achieving an accuracy of 100% on the CWRU dataset, while providing fast convergence, computational efficiency, and strong generalization capability. The proposed framework was further validated using an additional bearing dataset from the National Technical University ‘Kharkiv Polytechnic Institute’, achieving 99.97% classification accuracy and demonstrating strong robustness and generalization capability.</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 improved CNN–Transformer–ELM framework with residual shrinkage networks for bearing fault diagnosis

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Sujit Kumar,
  • Manish Kumar,
  • Bam Bahadur Sinha

摘要

Accurate fault diagnosis of rolling bearings is essential for ensuring the reliable operation of rotating machinery, particularly under noisy environments and complex operating conditions. To address this challenge, this paper presents a robust and noise-resilient bearing fault diagnosis framework that integrates advanced signal processing with hybrid deep learning techniques. Initially, raw vibration signals are decomposed using ensemble empirical mode decomposition (EEMD) to extract intrinsic mode functions that effectively capture fault-related characteristics under nonstationary conditions. Correlation coefficient-based analysis is then employed to identify the most informative components, thereby suppressing irrelevant and noise-dominated features. The selected features are normalized and encoded into two-channel Gramian angular field (GAF) representations, preserving temporal correlations and fault dynamics in a two-dimensional image domain. A convolutional neural network (CNN) embedded with an improved residual shrinkage network (IRSN) is subsequently utilized to perform adaptive noise suppression through learnable soft-thresholding while extracting discriminative spatial features. To further enhance representation capability and capture long-range dependencies under varying fault conditions, a Transformer encoder is incorporated after the CNN–IRSN stages, enabling effective modeling of global contextual information. The learned deep features are finally classified using an extreme learning machine (ELM), achieving an accuracy of 100% on the CWRU dataset, while providing fast convergence, computational efficiency, and strong generalization capability. The proposed framework was further validated using an additional bearing dataset from the National Technical University ‘Kharkiv Polytechnic Institute’, achieving 99.97% classification accuracy and demonstrating strong robustness and generalization capability.