Scholarlink

<p>Online surface roughness monitoring provides an effective way for ensuring the quality of thin-walled parts. During thin-wall milling, surface roughness is highly susceptible to the coupling effects of time-varying conditions, including workpiece thickness, process parameters, and cutting positions. However, these factors are usually neglected in traditional monitoring methods, thus leading to insufficient prediction accuracy. To address this issue, a method based on triple regularized multi-kernel autoencoder (TRMKAE) and most similar network transfer learning (MSNTL) is proposed to monitor surface roughness of thin-walled parts under variable milling conditions using acceleration, acoustic emission, and microphone signals. Firstly, multidomain features are extracted from the multi-sensory signals preprocessed by modulus maxima method. Subsequently, to highlight the nonlinear features closely correlated with the surface roughness of the thin-walled workpiece, a TRMKAE is established for feature fusion by mapping extracted multidomain features from the low-dimensional space to the high-dimensional space. Furthermore, a Jensen-Shannon divergence (JSD)-based MSNTL method is proposed to ensure the ability of TRMKAE in non-linear feature extraction under time-varying condition. The knowledge of source domain is easily transferred to target domain by minimizing JSD between two domains. Ultimately, the sensitive features are fed into the Kolmogorov–Arnold Network (KAN) for surface roughness monitoring. To verify the effectiveness of the proposed approach, a series of experiments on ball-end milling the PMMA thin-walled component are conducted, demonstrating superior accuracy and generality regardless of the operation conditions variation. The proposed online monitoring method achieves a Mean Relative Error (MRE) as low as 7.23% and a single prediction time of merely 11.7 ms, demonstrating higher prediction accuracy than traditional machine learning methods and faster processing speed compared to deep learning approaches.</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

Online surface roughness monitoring in thin-wall milling under time-varying conditions based on TRMKAE and MSNTL

  • 期刊论文

  • 发表日期：

    2026年2月14日

查看全文
  • Liang Sui,
  • Shichao Yan,
  • Yuwen Sun

摘要

Online surface roughness monitoring provides an effective way for ensuring the quality of thin-walled parts. During thin-wall milling, surface roughness is highly susceptible to the coupling effects of time-varying conditions, including workpiece thickness, process parameters, and cutting positions. However, these factors are usually neglected in traditional monitoring methods, thus leading to insufficient prediction accuracy. To address this issue, a method based on triple regularized multi-kernel autoencoder (TRMKAE) and most similar network transfer learning (MSNTL) is proposed to monitor surface roughness of thin-walled parts under variable milling conditions using acceleration, acoustic emission, and microphone signals. Firstly, multidomain features are extracted from the multi-sensory signals preprocessed by modulus maxima method. Subsequently, to highlight the nonlinear features closely correlated with the surface roughness of the thin-walled workpiece, a TRMKAE is established for feature fusion by mapping extracted multidomain features from the low-dimensional space to the high-dimensional space. Furthermore, a Jensen-Shannon divergence (JSD)-based MSNTL method is proposed to ensure the ability of TRMKAE in non-linear feature extraction under time-varying condition. The knowledge of source domain is easily transferred to target domain by minimizing JSD between two domains. Ultimately, the sensitive features are fed into the Kolmogorov–Arnold Network (KAN) for surface roughness monitoring. To verify the effectiveness of the proposed approach, a series of experiments on ball-end milling the PMMA thin-walled component are conducted, demonstrating superior accuracy and generality regardless of the operation conditions variation. The proposed online monitoring method achieves a Mean Relative Error (MRE) as low as 7.23% and a single prediction time of merely 11.7 ms, demonstrating higher prediction accuracy than traditional machine learning methods and faster processing speed compared to deep learning approaches.