Scholarlink

<p>Compound fault diagnosis in rotating machinery is challenging due to interactions among multiple components and the limited availability of labeled data. These challenges are further amplified under data imbalance conditions commonly encountered in practical manufacturing environments. To address this issue, this study proposes a data-driven framework that combines multi-sensor spectral feature fusion with Cycle Generative Adversarial Networks (CycleGANs) for compound fault diagnosis under data-scarce conditions. Multi-sensor time-domain signals acquired from acoustic and vibration sensors are first transformed into spectral representations and fused to capture cross-sensor fault characteristics. CycleGAN is then employed to generate realistic compound fault samples without requiring paired data, thereby reducing data scarcity and imbalance. A Convolutional Neural Network (CNN) is subsequently utilized for automated feature extraction and fault classification. Experimental validation on a rotating machinery fault simulator with seven fault categories demonstrates consistent diagnostic improvement across four imbalance severity levels (0.5, 0.25, 0.1, and 0.05). The proposed framework achieves average F1 scores of 0.90, 0.88, 0.86, and 0.85 at these imbalance levels, outperforming the best-performing baseline (Synthetic Minority Over-sampling combined CNN) by margins of 8%, 8%, 6%, and 11%, respectively. Statistical significance testing (two-sample t-test, <i>n</i> = 20, <i>p</i> &lt; 0.05) confirms that these improvements are robust across all tested conditions. The lower standard deviations observed for the proposed method (as low as 0.054) compared to baseline approaches further indicate improved diagnostic stability and reliability. The results demonstrate the effectiveness of the proposed approach for compound fault diagnosis and highlight its potential for predictive maintenance of rotating machinery in manufacturing systems.</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

Data-scarce compound fault diagnosis in rotating machinery using multi-sensor spectral fusion and CycleGAN-based data augmentation

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Ayantha Senanayaka,
  • Sungkwang Mun,
  • Abdullah Al Mamun,
  • Nayeon Lee

摘要

Compound fault diagnosis in rotating machinery is challenging due to interactions among multiple components and the limited availability of labeled data. These challenges are further amplified under data imbalance conditions commonly encountered in practical manufacturing environments. To address this issue, this study proposes a data-driven framework that combines multi-sensor spectral feature fusion with Cycle Generative Adversarial Networks (CycleGANs) for compound fault diagnosis under data-scarce conditions. Multi-sensor time-domain signals acquired from acoustic and vibration sensors are first transformed into spectral representations and fused to capture cross-sensor fault characteristics. CycleGAN is then employed to generate realistic compound fault samples without requiring paired data, thereby reducing data scarcity and imbalance. A Convolutional Neural Network (CNN) is subsequently utilized for automated feature extraction and fault classification. Experimental validation on a rotating machinery fault simulator with seven fault categories demonstrates consistent diagnostic improvement across four imbalance severity levels (0.5, 0.25, 0.1, and 0.05). The proposed framework achieves average F1 scores of 0.90, 0.88, 0.86, and 0.85 at these imbalance levels, outperforming the best-performing baseline (Synthetic Minority Over-sampling combined CNN) by margins of 8%, 8%, 6%, and 11%, respectively. Statistical significance testing (two-sample t-test, n = 20, p < 0.05) confirms that these improvements are robust across all tested conditions. The lower standard deviations observed for the proposed method (as low as 0.054) compared to baseline approaches further indicate improved diagnostic stability and reliability. The results demonstrate the effectiveness of the proposed approach for compound fault diagnosis and highlight its potential for predictive maintenance of rotating machinery in manufacturing systems.