Scholarlink

<p>Wind-turbine bearing vibrations are highly non-stationary and evolve slowly, so incipient faults are buried in noise and historical data contain very few labelled defect examples. Traditional data-hungry diagnostics therefore seldom apply. We introduce a dual-level information fusion network (DLIFN) that recasts the problem as a few-shot, self-supervised learning task. Time- and frequency-domain waveforms are concatenated into a two-channel image to exploit both transient impulses and periodic modulations; a lightweight, top-down architecture with lateral links and embedded multi-scale feature-fusion modules (MSFM) then aggregates cross-band, cross-scale descriptors. Before any target labels are seen, the network is pre-trained on abundant unlabeled source-domain records through a contrastive objective that incorporates a kurtosis-guided consistency term explicitly preserving the impulsive, heavy-tailed statistics that mark early defects. Fine-tuning with only 20% of target labels yields 99.3% mean diagnostic accuracy on four real bearing datasets, outperforming state-of-the-art methods by 1.9 percentage points while achieving a standard deviation below 0.5%. DLIFN achieves 94.4% accuracy when labelled data are reduced to 10% and generalises well across speeds, loads and bearing types, offering an interpretable, data-efficient solution for incipient fault detection in wind turbines and similar assets.</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

Dual-level information fusion network: A self-supervised time-frequency synergistic framework for few-shot incipient bearing fault diagnosis

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Qing Xu,
  • Xiangrong Song,
  • Yaobo Liu,
  • Qingchen Wang,
  • Dazhong Ma

摘要

Wind-turbine bearing vibrations are highly non-stationary and evolve slowly, so incipient faults are buried in noise and historical data contain very few labelled defect examples. Traditional data-hungry diagnostics therefore seldom apply. We introduce a dual-level information fusion network (DLIFN) that recasts the problem as a few-shot, self-supervised learning task. Time- and frequency-domain waveforms are concatenated into a two-channel image to exploit both transient impulses and periodic modulations; a lightweight, top-down architecture with lateral links and embedded multi-scale feature-fusion modules (MSFM) then aggregates cross-band, cross-scale descriptors. Before any target labels are seen, the network is pre-trained on abundant unlabeled source-domain records through a contrastive objective that incorporates a kurtosis-guided consistency term explicitly preserving the impulsive, heavy-tailed statistics that mark early defects. Fine-tuning with only 20% of target labels yields 99.3% mean diagnostic accuracy on four real bearing datasets, outperforming state-of-the-art methods by 1.9 percentage points while achieving a standard deviation below 0.5%. DLIFN achieves 94.4% accuracy when labelled data are reduced to 10% and generalises well across speeds, loads and bearing types, offering an interpretable, data-efficient solution for incipient fault detection in wind turbines and similar assets.