Scholarlink

Background <p>Fault diagnosis of industrial equipment with limited samples is challenging, as both traditional methods and conventional deep learning models require extensive labelled data. Existing Generative Adversarial Networks (GANs) for data augmentation primarily focus on time-domain signals, often neglecting frequency-domain feature consistency. Moreover, prevalent diagnosis models frequently suffer from feature redundancy, high computational complexity, and limited adaptability to complex fault patterns, which constrains their accuracy and generalisation in practical applications.</p> Purpose <p>This study proposes a novel serial fault diagnosis method that integrates a Fourier Transform Generative Adversarial Network (FTGAN) with a Self-Correcting Residual Separable Convolutional Neural Network (SCRSCNN). The objective is to overcome the small-sample problem and enhance diagnostic accuracy, stability, and generalisation.</p> Methods <p>The FTGAN leverages a Fourier transform to fuse time- and frequency-domain features, employing a dual-discriminator structure and a frequency consistency loss to generate high-quality augmented data. The SCRSCNN utilizes depthwise separable convolution, a self-correction mechanism, and a multi-branch fusion strategy for efficient feature extraction and dynamic adjustment. Experiments were conducted on two public datasets: the BJTURAO bogie and XJTU-SY bearing datasets. The proposed method's performance was benchmarked against existing approaches, including PSO-SVM, MSSCNN, and MTACNN. Ablation studies validated the contribution of each key component.</p> Results <p>On the BJTURAO dataset, the proposed method achieved a maximum diagnostic accuracy of 93.49% using 100 augmented samples. On the XJTU-SY dataset, it reached 100% accuracy under specific operating conditions. FTGAN-generated signals exhibited high consistency with real signals in both time and frequency domains, outperforming traditional GANs on metrics such as MSE and MMD. The SCRSCNN demonstrated superior stability, with a standard deviation of approximately 1% across tasks. Ablation experiments confirmed the significant role of each key component in both generation quality and diagnostic performance.</p> Conclusions <p>The integrated FTGAN-SCRSCNN method effectively mitigates the small-sample problem in industrial fault diagnosis by generating high-quality data and enabling efficient feature extraction. It surpasses existing methods in accuracy and stability, and its modular design facilitates practical industrial deployment. This approach provides a reliable solution for diagnosing faults in rotating machinery and demonstrates strong potential for application in complex real-world industrial environments.</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 New Fault Diagnosis Method Based on Fourier Transform Generative Adversarial Network and Self-correcting Residual Separable Convolutional Neural Network

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Jia Li,
  • Jiaxing Wang,
  • Hui Hu

摘要

Background

Fault diagnosis of industrial equipment with limited samples is challenging, as both traditional methods and conventional deep learning models require extensive labelled data. Existing Generative Adversarial Networks (GANs) for data augmentation primarily focus on time-domain signals, often neglecting frequency-domain feature consistency. Moreover, prevalent diagnosis models frequently suffer from feature redundancy, high computational complexity, and limited adaptability to complex fault patterns, which constrains their accuracy and generalisation in practical applications.

Purpose

This study proposes a novel serial fault diagnosis method that integrates a Fourier Transform Generative Adversarial Network (FTGAN) with a Self-Correcting Residual Separable Convolutional Neural Network (SCRSCNN). The objective is to overcome the small-sample problem and enhance diagnostic accuracy, stability, and generalisation.

Methods

The FTGAN leverages a Fourier transform to fuse time- and frequency-domain features, employing a dual-discriminator structure and a frequency consistency loss to generate high-quality augmented data. The SCRSCNN utilizes depthwise separable convolution, a self-correction mechanism, and a multi-branch fusion strategy for efficient feature extraction and dynamic adjustment. Experiments were conducted on two public datasets: the BJTURAO bogie and XJTU-SY bearing datasets. The proposed method's performance was benchmarked against existing approaches, including PSO-SVM, MSSCNN, and MTACNN. Ablation studies validated the contribution of each key component.

Results

On the BJTURAO dataset, the proposed method achieved a maximum diagnostic accuracy of 93.49% using 100 augmented samples. On the XJTU-SY dataset, it reached 100% accuracy under specific operating conditions. FTGAN-generated signals exhibited high consistency with real signals in both time and frequency domains, outperforming traditional GANs on metrics such as MSE and MMD. The SCRSCNN demonstrated superior stability, with a standard deviation of approximately 1% across tasks. Ablation experiments confirmed the significant role of each key component in both generation quality and diagnostic performance.

Conclusions

The integrated FTGAN-SCRSCNN method effectively mitigates the small-sample problem in industrial fault diagnosis by generating high-quality data and enabling efficient feature extraction. It surpasses existing methods in accuracy and stability, and its modular design facilitates practical industrial deployment. This approach provides a reliable solution for diagnosing faults in rotating machinery and demonstrates strong potential for application in complex real-world industrial environments.