Scholarlink

<p>Accurate and timely bearing fault diagnosis through the analysis of vibration signals is conducive to improving the maintenance efficiency and ensuring the overall safety and reliability of aero-engines. However, existing denoising methods cannot effectively maintain the integrity of operating signals with random noise removed, influencing the stability of later fault diagnosis. And due to a certain degree of signal amplitude and frequency changing in different working conditions, some diagnosis models cannot maintain the capability of distinguishing potential faults well. To address the above problems, we propose a CEAE-CGIVT hybrid model for aero-engine bearing fault diagnosis. CEAE-CGIVT consists of a columnar expansion convolutional autoencoder (CEAE), ResNet10 and cross-granularity vision transformer (CGIVT). Specifically, CEAE respectively conducts the dilated convolution after encoding and decoding to more fully dilute random noise points and preserve potential signal features. And it employs a columnar structure to avoid losing the detail information. Then, feed clean signals into shallow ResNet10 network to initially enrich and transfer local features. Furthermore, ADPE module in CGIVT uses two parallel double-telescopic multi-head product self-attention mechanisms (Dt-PSAM) to synchronously explore and enhance mixed granularity features at dual scales, improving the diagnostic effect of potential bearing faults under extreme working conditions. Experiments on DIRG bearing datasets and HUST bearing datasets show that CEAE-CGIVT outperforms other state-of-the-art methods and demonstrates good adaptability and robustness.</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

CEAE-CGIVT: cross-granularity interactive vision transformer with columnar expansion denoising module for aero-engine bearing fault diagnosis

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Lihao Zhou,
  • Huawei Wang,
  • Liang Zhou

摘要

Accurate and timely bearing fault diagnosis through the analysis of vibration signals is conducive to improving the maintenance efficiency and ensuring the overall safety and reliability of aero-engines. However, existing denoising methods cannot effectively maintain the integrity of operating signals with random noise removed, influencing the stability of later fault diagnosis. And due to a certain degree of signal amplitude and frequency changing in different working conditions, some diagnosis models cannot maintain the capability of distinguishing potential faults well. To address the above problems, we propose a CEAE-CGIVT hybrid model for aero-engine bearing fault diagnosis. CEAE-CGIVT consists of a columnar expansion convolutional autoencoder (CEAE), ResNet10 and cross-granularity vision transformer (CGIVT). Specifically, CEAE respectively conducts the dilated convolution after encoding and decoding to more fully dilute random noise points and preserve potential signal features. And it employs a columnar structure to avoid losing the detail information. Then, feed clean signals into shallow ResNet10 network to initially enrich and transfer local features. Furthermore, ADPE module in CGIVT uses two parallel double-telescopic multi-head product self-attention mechanisms (Dt-PSAM) to synchronously explore and enhance mixed granularity features at dual scales, improving the diagnostic effect of potential bearing faults under extreme working conditions. Experiments on DIRG bearing datasets and HUST bearing datasets show that CEAE-CGIVT outperforms other state-of-the-art methods and demonstrates good adaptability and robustness.