Scholarlink

<p>As a power source in the hydraulic system, the performance of the double-acting vane pump directly determines the reliability and safety of the overall system operation. In order to meet the needs of non-invasive measurement in some scenarios and solve the deficiency of diagnostic accuracy caused by signal type constraints, a framework of triple-attention mechanism combined with residual network is proposed. Firstly, the one-dimensional triaxial vibration signal is converted into a two-dimensional RGB grayscale image by using the Gram angle difference field (GADF). Then the data level fusion and the automatic allocation of the attention weight of the three-channel data are realized by the triple-attention mechanism. Finally, the fault diagnosis is completed by the deep residual networks (ResNet). By building the vane pump fault simulation test bench, the original data under the five labels of normal, the inner surface wear of stator, the blade crack, the end face wear of rotor and the wear of distribution plate were collected. The TA-ResNet model was verified and compared with the basic RGB fusion method and Vision Transformer (VIT) model. The robustness of the TA-ResNet model is to be verified by adding Gaussian white noise with different signal-to-noise ratios (SNRs) to the original data. The results show that the TA-ResNet model is superior to the RGB fusion method and the VIT model in terms of comprehensive performance. The accuracy is 98.50%, the precision is 98.50%, the recall is 98.52%, and the <i>F</i><sub>1</sub> value is 98.51%. The accuracy of the TA-ResNet model is 96.74% at the SNR of 7dB. It shows that the proposed TA-ResNet model has stronger ability in channel attention allocation than the VIT model, and proves the effectiveness and robustness of the model.</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

Fault diagnosis method of double-acting vane pump based on three-axis vibration signal fusion

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Rui Guo,
  • Hao Zhang,
  • Jiaqi Zhang,
  • Wenwen Niu,
  • Jingyi Zhao,
  • Jianwei Wang,
  • Wei Cai

摘要

As a power source in the hydraulic system, the performance of the double-acting vane pump directly determines the reliability and safety of the overall system operation. In order to meet the needs of non-invasive measurement in some scenarios and solve the deficiency of diagnostic accuracy caused by signal type constraints, a framework of triple-attention mechanism combined with residual network is proposed. Firstly, the one-dimensional triaxial vibration signal is converted into a two-dimensional RGB grayscale image by using the Gram angle difference field (GADF). Then the data level fusion and the automatic allocation of the attention weight of the three-channel data are realized by the triple-attention mechanism. Finally, the fault diagnosis is completed by the deep residual networks (ResNet). By building the vane pump fault simulation test bench, the original data under the five labels of normal, the inner surface wear of stator, the blade crack, the end face wear of rotor and the wear of distribution plate were collected. The TA-ResNet model was verified and compared with the basic RGB fusion method and Vision Transformer (VIT) model. The robustness of the TA-ResNet model is to be verified by adding Gaussian white noise with different signal-to-noise ratios (SNRs) to the original data. The results show that the TA-ResNet model is superior to the RGB fusion method and the VIT model in terms of comprehensive performance. The accuracy is 98.50%, the precision is 98.50%, the recall is 98.52%, and the F1 value is 98.51%. The accuracy of the TA-ResNet model is 96.74% at the SNR of 7dB. It shows that the proposed TA-ResNet model has stronger ability in channel attention allocation than the VIT model, and proves the effectiveness and robustness of the model.