Scholarlink

<p>Mechanical transmission systems often exhibit feature distribution shifts under different operating conditions, limiting the generalizability of supervised fault diagnosis methods. This paper proposes an unsupervised transfer learning method that combines frequency-domain Markov Transition Field (MTF) images with an improved Deep Sub-domain Adaptation (DSDA) network to achieve multi-component cross-domain fault diagnosis. In this method, vibration signals are first converted into the frequency domain by fast Fourier transform (FFT) to generate feature-enhanced MTF images. The improved DSDA network then employs a newly designed Multi-scale Residual Network (MSResNet) to extract transferable multi-level features and a Hybrid Distance Metric (HDM) strategy for the joint alignment of both marginal and fine-grained conditional distributions between domains. Experimental data from 11 typical faults in bearings, rotors, and gears were collected on a laboratory test rig and used to construct 12 transfer tasks covering various speed and load configurations. The results show that the proposed method achieves average diagnostic accuracy of 90.48% and 97.48% in cross-speed and cross-load tasks, respectively, validating its excellent cross-domain fault diagnosis capability.</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

Cross-domain multi-component fault diagnosis using frequency-domain image transformation and joint alignment

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Hongwei Fan,
  • Jie Li,
  • Xiangang Cao,
  • Xuhui Zhang

摘要

Mechanical transmission systems often exhibit feature distribution shifts under different operating conditions, limiting the generalizability of supervised fault diagnosis methods. This paper proposes an unsupervised transfer learning method that combines frequency-domain Markov Transition Field (MTF) images with an improved Deep Sub-domain Adaptation (DSDA) network to achieve multi-component cross-domain fault diagnosis. In this method, vibration signals are first converted into the frequency domain by fast Fourier transform (FFT) to generate feature-enhanced MTF images. The improved DSDA network then employs a newly designed Multi-scale Residual Network (MSResNet) to extract transferable multi-level features and a Hybrid Distance Metric (HDM) strategy for the joint alignment of both marginal and fine-grained conditional distributions between domains. Experimental data from 11 typical faults in bearings, rotors, and gears were collected on a laboratory test rig and used to construct 12 transfer tasks covering various speed and load configurations. The results show that the proposed method achieves average diagnostic accuracy of 90.48% and 97.48% in cross-speed and cross-load tasks, respectively, validating its excellent cross-domain fault diagnosis capability.