Scholarlink

<p>In dynamic environments such as industrial settings, complex noise interference often leads to unstable performance in the diagnosis of rotating machinery faults. This paper proposes a combined Mel-scale time–frequency processing multi-scale convolutional neural network (Mel-MSCNN) to improve the accuracy and robustness of fault diagnosis by combining time-series data processing and network structure optimization. The method applies Mel-scale nonlinear time–frequency processing to filter multiple time-series data sets. It converts them into Mel time–frequency grayscale images, which are then stacked by channels to generate multi-dimensional tensor features. To address high-dimensional input features, an attention mechanism that combines spatial, channel, and coordinate dimensions is introduced, highlighting important spatial location information, key channel feature representations, and coordinate information. Additionally, multiple bidirectional spatial pyramid convolution modules are embedded in the model to capture and integrate multi-scale feature information, enhancing model performance and generalization capability. The main structure employs skip connections, allowing gradients to be directly transferred from shallow to deep layers, ensuring stability and convergence speed during model training. Testing on public bearing and gear datasets showed classification accuracies exceeding 98%, with the multiply-accumulate operations (MACs) and parameters being 0.87G and 5.40M, respectively. The experimental results demonstrate that the proposed method performs well in mechanical fault diagnosis.</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 noise-robust fault diagnosis method for rotating machinery in dynamic environments based on multi-scale convolution

  • 期刊论文

  • 发表日期：

    2025年9月1日

查看全文
  • Jianbo Guo,
  • Shuai Wang,
  • Wei Guan,
  • Guoqiang Wang,
  • Zhengbin Liu,
  • Shuwei Wu

摘要

In dynamic environments such as industrial settings, complex noise interference often leads to unstable performance in the diagnosis of rotating machinery faults. This paper proposes a combined Mel-scale time–frequency processing multi-scale convolutional neural network (Mel-MSCNN) to improve the accuracy and robustness of fault diagnosis by combining time-series data processing and network structure optimization. The method applies Mel-scale nonlinear time–frequency processing to filter multiple time-series data sets. It converts them into Mel time–frequency grayscale images, which are then stacked by channels to generate multi-dimensional tensor features. To address high-dimensional input features, an attention mechanism that combines spatial, channel, and coordinate dimensions is introduced, highlighting important spatial location information, key channel feature representations, and coordinate information. Additionally, multiple bidirectional spatial pyramid convolution modules are embedded in the model to capture and integrate multi-scale feature information, enhancing model performance and generalization capability. The main structure employs skip connections, allowing gradients to be directly transferred from shallow to deep layers, ensuring stability and convergence speed during model training. Testing on public bearing and gear datasets showed classification accuracies exceeding 98%, with the multiply-accumulate operations (MACs) and parameters being 0.87G and 5.40M, respectively. The experimental results demonstrate that the proposed method performs well in mechanical fault diagnosis.