Scholarlink

Purpose <p>Conventional supervised learning methods for fault diagnosis require extensive labeled data, which is often unavailable due to the sparsity and high cost of labeling fault signals. This study aims to develop a self-supervised learning approach that can effectively extract discriminative features from unlabeled vibration data, enabling accurate fault diagnosis with minimal labeled samples.</p> Methods <p>A time–frequency residual self-supervised network (TFRSSN) is proposed, based on an improved MoBY framework. The Transformer backbone in the original MoBY architecture is replaced with a 1D ResNet-18 to accommodate one-dimensional vibration signals and reduce model complexity. Paired time–frequency representations are generated as input for contrastive pretraining. A dynamic queue is employed to store negative samples for enhancing contrastive learning. Several data augmentation techniques specifically designed for fault signals are developed, and the optimal combination is determined through ablation experiments. After self-supervised pretraining, the network can be adapted to specific diagnostic tasks by freezing or fine-tuning selected layers.</p> Results <p>The proposed TFRSSN is evaluated on two benchmark bearing fault datasets: SDUST and PU. Experimental results show that the model achieves competitive diagnostic accuracy even when fine-tuned with only 1% of labeled data. The integration of contrastive learning, time–frequency representations, and tailored augmentation strategies significantly improves feature extraction performance.</p> Conclusion <p>The TFRSSN model demonstrates strong potential for real-world fault diagnosis under limited labeled data conditions. Its effectiveness in learning robust representations from unlabeled signals makes it a promising solution for industrial applications where labeled data is scarce.</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

Time–Frequency Residual Self-Supervised Network for Bearing Fault Diagnosis with Limited Labeled Data

  • 期刊论文

  • 发表日期：

    2025年7月28日

查看全文
  • Zhikang Fan,
  • Zongzhen Zhang,
  • Jinrui Wang,
  • Huaiqian Bao,
  • Baokun Han,
  • Wenqi Wang,
  • Rongkang Ge,
  • Junji Ma

摘要

Purpose

Conventional supervised learning methods for fault diagnosis require extensive labeled data, which is often unavailable due to the sparsity and high cost of labeling fault signals. This study aims to develop a self-supervised learning approach that can effectively extract discriminative features from unlabeled vibration data, enabling accurate fault diagnosis with minimal labeled samples.

Methods

A time–frequency residual self-supervised network (TFRSSN) is proposed, based on an improved MoBY framework. The Transformer backbone in the original MoBY architecture is replaced with a 1D ResNet-18 to accommodate one-dimensional vibration signals and reduce model complexity. Paired time–frequency representations are generated as input for contrastive pretraining. A dynamic queue is employed to store negative samples for enhancing contrastive learning. Several data augmentation techniques specifically designed for fault signals are developed, and the optimal combination is determined through ablation experiments. After self-supervised pretraining, the network can be adapted to specific diagnostic tasks by freezing or fine-tuning selected layers.

Results

The proposed TFRSSN is evaluated on two benchmark bearing fault datasets: SDUST and PU. Experimental results show that the model achieves competitive diagnostic accuracy even when fine-tuned with only 1% of labeled data. The integration of contrastive learning, time–frequency representations, and tailored augmentation strategies significantly improves feature extraction performance.

Conclusion

The TFRSSN model demonstrates strong potential for real-world fault diagnosis under limited labeled data conditions. Its effectiveness in learning robust representations from unlabeled signals makes it a promising solution for industrial applications where labeled data is scarce.