Scholarlink

<p>Rotary equipment, prevalent in industries like production, aeronautics, conveyance, and healthcare apparatus, is crucial since malfunctions could compromise the integrity of goods and potentially pose a threat to human safety. Traditional methods for feature extraction are complex and slow to train, making it difficult to achieve high-accuracy models with limited datasets in variable environments. The manuscript introduces a fusion framework based on multiple deep learning networks utilizing transfer learning for detecting faults in rotating equipment operating under varied conditions. Initially, we standardize the original vibration signals from the rotating frequency, drive end, and fan end, and encode them into the three channels of an image to obtain a feature map with complete information and its label. Then, the feature map is preprocessed, the feature extraction layers of WideResNet, ResNeSt, and ResNet152 are frozen, and the model is fine-tuned and fused across multiple networks before being input into the multi-network fusion transfer model. The methodology adopted here imports parameters from the originating structures and amalgamates the capabilities of various networks, thereby improving the model’s ability to diagnose errors accurately and boosting its broad applicability when data is scarce. Moreover, it notably diminishes the duration needed to train the model. Validation of the suggested model occurred through the use of bearing data sourced from Case Western Reserve University, where it attained a perfect score in distinguishing between two classes and a 99.7 % success rate in differentiating among ten classes across a range of operating environments. It demonstrated superior stability across various conditions, outperforming other fault diagnosis models.</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

Few-shot transfer learning on multiple deep networks for bearing fault diagnosis under variable working conditions

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Junyan Wang,
  • Mingqiang Yang,
  • Mingbao Yang

摘要

Rotary equipment, prevalent in industries like production, aeronautics, conveyance, and healthcare apparatus, is crucial since malfunctions could compromise the integrity of goods and potentially pose a threat to human safety. Traditional methods for feature extraction are complex and slow to train, making it difficult to achieve high-accuracy models with limited datasets in variable environments. The manuscript introduces a fusion framework based on multiple deep learning networks utilizing transfer learning for detecting faults in rotating equipment operating under varied conditions. Initially, we standardize the original vibration signals from the rotating frequency, drive end, and fan end, and encode them into the three channels of an image to obtain a feature map with complete information and its label. Then, the feature map is preprocessed, the feature extraction layers of WideResNet, ResNeSt, and ResNet152 are frozen, and the model is fine-tuned and fused across multiple networks before being input into the multi-network fusion transfer model. The methodology adopted here imports parameters from the originating structures and amalgamates the capabilities of various networks, thereby improving the model’s ability to diagnose errors accurately and boosting its broad applicability when data is scarce. Moreover, it notably diminishes the duration needed to train the model. Validation of the suggested model occurred through the use of bearing data sourced from Case Western Reserve University, where it attained a perfect score in distinguishing between two classes and a 99.7 % success rate in differentiating among ten classes across a range of operating environments. It demonstrated superior stability across various conditions, outperforming other fault diagnosis models.