Scholarlink

<p>To overcome the limitations of conventional experimental and analytical approaches that depend on offline inspection and lack real-time process optimization capabilities, this study developed the Relative Surface Roughness Convolutional Neural Network-Long Short-Term Memory (R-CNN-LSTM) model for predicting surface roughness in 6061 aluminum alloy components using 3D force time-series signals. Based on the data characteristics, we systematically optimized the hyperparameters of the CNN component, and implemented comparative models, including R-CNN, R-LSTM and CNN-LSTM specifically for the semi-finish polishing roughness range. Furthermore, feature visualization was employed to analyze the feature extraction process of R-CNN-LSTM, providing insights into its learning mechanism. Comprehensive evaluation of test loss curves, prediction accuracy, and cross-validation results demonstrates that the proposed R-CNN-LSTM architecture achieves superior convergence speed and prediction accuracy within the semi-finish polishing roughness range, while exhibiting optimal generalization capability and stability.</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

Deep learning-based prediction of relative surface roughness using dynamic force signals

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Jiahao He,
  • Haibin Huang,
  • Jiaming Xiong,
  • Zihao Ye

摘要

To overcome the limitations of conventional experimental and analytical approaches that depend on offline inspection and lack real-time process optimization capabilities, this study developed the Relative Surface Roughness Convolutional Neural Network-Long Short-Term Memory (R-CNN-LSTM) model for predicting surface roughness in 6061 aluminum alloy components using 3D force time-series signals. Based on the data characteristics, we systematically optimized the hyperparameters of the CNN component, and implemented comparative models, including R-CNN, R-LSTM and CNN-LSTM specifically for the semi-finish polishing roughness range. Furthermore, feature visualization was employed to analyze the feature extraction process of R-CNN-LSTM, providing insights into its learning mechanism. Comprehensive evaluation of test loss curves, prediction accuracy, and cross-validation results demonstrates that the proposed R-CNN-LSTM architecture achieves superior convergence speed and prediction accuracy within the semi-finish polishing roughness range, while exhibiting optimal generalization capability and stability.