Scholarlink

<p>Transitioning from preventive to predictive maintenance is critical for enhancing the economic performance of wind farms. This study proposes a dual-source data-driven diagnostic framework for wind turbine generator bearings, leveraging industrial-grade vibration sensors and smartphone microphones to capture mechanical health data. This study investigates the efficacy of three distinct deep learning architectures–convolutional neural networks, long short-term memory networks, and the transformer encoder–for fault diagnosis. A dataset comprising 16 fault types, including inner race, outer race, rolling element, and cage defects of varying severity, was constructed to train and evaluate the models. The experimental analysis compared the classification accuracies of vibration and acoustic signals across different network configurations, such as the number of convolutional blocks, memory units, and attention heads. The results indicate that although vibration signals generally yield higher accuracy, deep learning models can effectively utilize smartphone-acquired acoustic data for fault identification. Specifically, the Transformer model achieved a peak accuracy of 95.1 ± 1.1% using acoustic signal, demonstrating the potential of attention mechanisms for capturing complex fault features. This study validates the feasibility of employing dual-source data and neural networks for intelligent 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

Data-driven fault diagnosis of wind turbine generator bearings via heterogeneous vibration and acoustic sensing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Xiaomeng Li,
  • Xiaoxia Song

摘要

Transitioning from preventive to predictive maintenance is critical for enhancing the economic performance of wind farms. This study proposes a dual-source data-driven diagnostic framework for wind turbine generator bearings, leveraging industrial-grade vibration sensors and smartphone microphones to capture mechanical health data. This study investigates the efficacy of three distinct deep learning architectures–convolutional neural networks, long short-term memory networks, and the transformer encoder–for fault diagnosis. A dataset comprising 16 fault types, including inner race, outer race, rolling element, and cage defects of varying severity, was constructed to train and evaluate the models. The experimental analysis compared the classification accuracies of vibration and acoustic signals across different network configurations, such as the number of convolutional blocks, memory units, and attention heads. The results indicate that although vibration signals generally yield higher accuracy, deep learning models can effectively utilize smartphone-acquired acoustic data for fault identification. Specifically, the Transformer model achieved a peak accuracy of 95.1 ± 1.1% using acoustic signal, demonstrating the potential of attention mechanisms for capturing complex fault features. This study validates the feasibility of employing dual-source data and neural networks for intelligent fault diagnosis.