Scholarlink

<p>Accurate tool wear monitoring is pivotal for ensuring machining precision in intelligent manufacturing systems. To overcome the limitations of traditional single-sensor approaches in exploiting multi-source heterogeneous signals, this study proposes a dual-channel CNN-LSTM framework integrating time-frequency and temporal feature fusion. EMD-HHT was used to generate intrinsic mode functions (IMFs), from which Hilbert spectral analysis was performed to construct 2D time-frequency representations (TFRs) encapsulating instantaneous amplitude-frequency characteristics. These TFRs were processed through a Convolutional Neural Network (CNN) combining residual blocks and multi-head attention mechanisms to extract hierarchical spatial features. In parallel, an attention-enhanced Long Short-Term Memory (LSTM) network was used to capture long-range temporal dependencies from denoised signal sequences. The features extracted from both channels were concatenated and fed into a fully connected layer to perform the final classification task. Experimental validation on a five-axis machining dataset, using a Leave-One-Cutter-Out Cross-Validation (LOCO-CV) strategy, shows that the proposed method achieves an average accuracy of 94.28%, outperforming conventional methods in generalization 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

A Hybrid CNN-LSTM method combined with EMD-HHT feature extraction for tool wear monitoring

  • 期刊论文

  • 发表日期：

    2026年5月9日

查看全文
  • Yang Dai,
  • Shi-xian Xuan,
  • Bing-can LI,
  • Yu-heng Dai,
  • Sheng-peng Luo,
  • Zuo Jiang

摘要

Accurate tool wear monitoring is pivotal for ensuring machining precision in intelligent manufacturing systems. To overcome the limitations of traditional single-sensor approaches in exploiting multi-source heterogeneous signals, this study proposes a dual-channel CNN-LSTM framework integrating time-frequency and temporal feature fusion. EMD-HHT was used to generate intrinsic mode functions (IMFs), from which Hilbert spectral analysis was performed to construct 2D time-frequency representations (TFRs) encapsulating instantaneous amplitude-frequency characteristics. These TFRs were processed through a Convolutional Neural Network (CNN) combining residual blocks and multi-head attention mechanisms to extract hierarchical spatial features. In parallel, an attention-enhanced Long Short-Term Memory (LSTM) network was used to capture long-range temporal dependencies from denoised signal sequences. The features extracted from both channels were concatenated and fed into a fully connected layer to perform the final classification task. Experimental validation on a five-axis machining dataset, using a Leave-One-Cutter-Out Cross-Validation (LOCO-CV) strategy, shows that the proposed method achieves an average accuracy of 94.28%, outperforming conventional methods in generalization capability.