Scholarlink

<p>The accurate diagnosis of faults in rotating machinery within power-generation systems is critically challenged by severe class imbalance and scarce fault samples, which substantially degrade the performance of conventional deep learning models. To address these challenges, this paper proposes a three-component diagnostic framework. First, a composite data augmentation strategy combining Gaussian noise injection, temporal shifting, and random amplitude scaling is employed to expand minority fault samples and restore class balance in the training set. Second, a Dilated Residual Network (DilatedResNet) is designed for multi-scale feature extraction, where dilated convolutions with expansion factors of 1, 2, and 4 simultaneously capture fine-grained impulse signatures and broad periodic patterns without resolution loss caused by pooling, while residual connections maintain stable gradient propagation. Third, an Extreme Learning Machine (ELM) is investigated as a lightweight alternative to stacked fully connected (FC) layers, computing output weights analytically through matrix inversion to reduce training complexity and mitigate overfitting. The proposed DilatedResNet with data augmentation achieves a diagnostic accuracy of 99.72% ± 0.63% over ten independent runs on the CWRU bearing dataset with a approximately 24:1 imbalance ratio, representing a 1.40 percentage point improvement over the baseline. The ELM classifier provides a practical lightweight alternative with competitive accuracy of 99.07% ± 0.88% and a 205 × reduction in training time compared to the FC-based classifier. The ablation study systematically validates the contribution of each proposed component. Ross-condition generalization experiments on the CWRU dataset demonstrate that the proposed method maintains an average accuracy of 91.45% when transferring from the training condition (0 hp) to unseen load conditions (1 hp, 2 hp, and 3 hp), outperforming the baseline by 10.54 percentage points.</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 fault diagnosis method for power generation rotating machinery based on small-sample augmentation and dilated residual network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月9日

查看全文
  • Lipeng Wang,
  • Junchao Ge

摘要

The accurate diagnosis of faults in rotating machinery within power-generation systems is critically challenged by severe class imbalance and scarce fault samples, which substantially degrade the performance of conventional deep learning models. To address these challenges, this paper proposes a three-component diagnostic framework. First, a composite data augmentation strategy combining Gaussian noise injection, temporal shifting, and random amplitude scaling is employed to expand minority fault samples and restore class balance in the training set. Second, a Dilated Residual Network (DilatedResNet) is designed for multi-scale feature extraction, where dilated convolutions with expansion factors of 1, 2, and 4 simultaneously capture fine-grained impulse signatures and broad periodic patterns without resolution loss caused by pooling, while residual connections maintain stable gradient propagation. Third, an Extreme Learning Machine (ELM) is investigated as a lightweight alternative to stacked fully connected (FC) layers, computing output weights analytically through matrix inversion to reduce training complexity and mitigate overfitting. The proposed DilatedResNet with data augmentation achieves a diagnostic accuracy of 99.72% ± 0.63% over ten independent runs on the CWRU bearing dataset with a approximately 24:1 imbalance ratio, representing a 1.40 percentage point improvement over the baseline. The ELM classifier provides a practical lightweight alternative with competitive accuracy of 99.07% ± 0.88% and a 205 × reduction in training time compared to the FC-based classifier. The ablation study systematically validates the contribution of each proposed component. Ross-condition generalization experiments on the CWRU dataset demonstrate that the proposed method maintains an average accuracy of 91.45% when transferring from the training condition (0 hp) to unseen load conditions (1 hp, 2 hp, and 3 hp), outperforming the baseline by 10.54 percentage points.