Scholarlink

<p>Supervisory control and data acquisition (SCADA) systems on wind turbines generate extensive sensory data essential for fault diagnosis. SCADA data consist of multivariate time-series with complex temporal correlations within each sensor variable and spatial correlations among different variables. However, noise from environmental variability, sensor errors, and irregular time-series data can distort these patterns, making feature extraction challenging and reducing reliability in fault diagnosis. To address these issues, this study presents a novel model integrating residual neural networks with bi-level Gaussian noise augmentation and time-aware long short-term memory networks (BiGN-ResNet-T-LSTM), leveraging fixed-scale kernels within ResNet layers. For mitigating SCADA data imbalance caused by abundant normal data and scarce fault data, we employ the synthetic minority oversampling technique and compare its performance with cross-entropy and focal loss functions. Validated on two distinct datasets for fault classification and parameter prediction, the BiGN-ResNet-T-LSTM framework significantly outperforms the existing methods. Additionally, the proposed model is deployed in cloud environments using Amazon SageMaker, enabling secure and scalable online condition monitoring and real-time fault diagnosis.</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 Fixed Kernel and Cloud-Integrated Deep Learning Model for Wind Turbine Fault Diagnosis with Imbalanced SCADA Data

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • V Siva Brahmaiah Rama,
  • Jung-Min Yang

摘要

Supervisory control and data acquisition (SCADA) systems on wind turbines generate extensive sensory data essential for fault diagnosis. SCADA data consist of multivariate time-series with complex temporal correlations within each sensor variable and spatial correlations among different variables. However, noise from environmental variability, sensor errors, and irregular time-series data can distort these patterns, making feature extraction challenging and reducing reliability in fault diagnosis. To address these issues, this study presents a novel model integrating residual neural networks with bi-level Gaussian noise augmentation and time-aware long short-term memory networks (BiGN-ResNet-T-LSTM), leveraging fixed-scale kernels within ResNet layers. For mitigating SCADA data imbalance caused by abundant normal data and scarce fault data, we employ the synthetic minority oversampling technique and compare its performance with cross-entropy and focal loss functions. Validated on two distinct datasets for fault classification and parameter prediction, the BiGN-ResNet-T-LSTM framework significantly outperforms the existing methods. Additionally, the proposed model is deployed in cloud environments using Amazon SageMaker, enabling secure and scalable online condition monitoring and real-time fault diagnosis.