Scholarlink

<p>Rotating machinery serves as a cornerstone of modern industrial systems, and accurate fault diagnosis is essential for ensuring operational safety and efficiency. However, existing diagnostic models often struggle to balance local feature extraction and global dependency modeling, limiting their ability to identify complex fault patterns. To address this challenge, this paper proposes a novel CNN-Transformer hybrid architecture enhanced with a Multi-scale Adaptive Cross-Attention Mechanism (MACAM). The framework achieves heterogeneous integration of convolutional neural networks’ local pattern recognition capabilities and Transformers’ global context modeling strengths through two innovative components: (1) a dynamic weight adaptation module (α-module) for time–frequency signal fusion optimization, and (2) a feature contribution calibration mechanism (β-module) enabling domain-specific feature re-calibration. Comprehensive experimental validation on the Jiangnan University bearing dataset demonstrates the model’s superior performance in 12-class fault classification tasks. Quantitative results reveal that our model achieves state-of-the-art performance with a leading average accuracy of 99.1%, surpassing conventional CNN-Attention (94.23%), CNN-LSTM (96.91%), and baseline CNN-Transformer (95.68%) architectures by margins of 4.87%, 2.19%, and 3.42%, respectively. Notably, it achieves a 12.7-percentage-point improvement in accuracy for fault categories with overlapping spectral characteristics, demonstrating enhanced discriminative capability in complex diagnostic scenarios. The proposed adaptive attention paradigm also exhibits exceptional computational efficiency, achieving perfect testing accuracy within 40 epochs while reducing training cycles by 20%—yielding convergence acceleration factors ranging from 1.42 × to 1.85 × compared to conventional approaches. These results substantiate the model’s advantages in robustness and generalization capability, providing an effective framework for intelligent fault diagnosis in rotating machinery through adaptive multimodal feature fusion.</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

Research on fault diagnosis method based on improved CNN-transformer for rotating equipment

  • 期刊论文

  • 发表日期：

    2025年11月23日

查看全文
  • Dali Hou,
  • XueYan Ma

摘要

Rotating machinery serves as a cornerstone of modern industrial systems, and accurate fault diagnosis is essential for ensuring operational safety and efficiency. However, existing diagnostic models often struggle to balance local feature extraction and global dependency modeling, limiting their ability to identify complex fault patterns. To address this challenge, this paper proposes a novel CNN-Transformer hybrid architecture enhanced with a Multi-scale Adaptive Cross-Attention Mechanism (MACAM). The framework achieves heterogeneous integration of convolutional neural networks’ local pattern recognition capabilities and Transformers’ global context modeling strengths through two innovative components: (1) a dynamic weight adaptation module (α-module) for time–frequency signal fusion optimization, and (2) a feature contribution calibration mechanism (β-module) enabling domain-specific feature re-calibration. Comprehensive experimental validation on the Jiangnan University bearing dataset demonstrates the model’s superior performance in 12-class fault classification tasks. Quantitative results reveal that our model achieves state-of-the-art performance with a leading average accuracy of 99.1%, surpassing conventional CNN-Attention (94.23%), CNN-LSTM (96.91%), and baseline CNN-Transformer (95.68%) architectures by margins of 4.87%, 2.19%, and 3.42%, respectively. Notably, it achieves a 12.7-percentage-point improvement in accuracy for fault categories with overlapping spectral characteristics, demonstrating enhanced discriminative capability in complex diagnostic scenarios. The proposed adaptive attention paradigm also exhibits exceptional computational efficiency, achieving perfect testing accuracy within 40 epochs while reducing training cycles by 20%—yielding convergence acceleration factors ranging from 1.42 × to 1.85 × compared to conventional approaches. These results substantiate the model’s advantages in robustness and generalization capability, providing an effective framework for intelligent fault diagnosis in rotating machinery through adaptive multimodal feature fusion.