Scholarlink

<p>In high-precision manufacturing, surface roughness inspection is crucial for quality assurance; however, conventional contact-based methods are slow and costly. This study presents a physics-guided artificial intelligence framework for non-contact roughness prediction, integrating frequency-domain feature engineering with lightweight neural network regression. A novel circular–sector decomposition of the power spectral density (PSD) is proposed to extract roughness-sensitive texture descriptors that encode both frequency energy distribution and directional anisotropy. These handcrafted, physics-informed features reduce data demands and improve interpretability compared to black-box deep learning methods. A hybrid backpropagation neural network (BP NN), designed as a nonlinear function approximator, predicts surface roughness with &lt; 7% error, outperforming conventional statistical methods and run-length matrix-based models. Additionally, a multimodal extension incorporating vibration and acoustic signals demonstrates robustness to machining variations, highlighting the system’s potential for real-time, edge-deployable intelligent quality inspection. This work indicates that integrating physics-driven spectral analysis with machine learning creates a scalable, explainable, and high-accuracy AI solution for Industry 4.0 manufacturing.</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

Intelligent detection technology of machined surface materials based on image nonlinear processing algorithm

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月21日

查看全文
  • Aiwu Tang,
  • Yi Chen

摘要

In high-precision manufacturing, surface roughness inspection is crucial for quality assurance; however, conventional contact-based methods are slow and costly. This study presents a physics-guided artificial intelligence framework for non-contact roughness prediction, integrating frequency-domain feature engineering with lightweight neural network regression. A novel circular–sector decomposition of the power spectral density (PSD) is proposed to extract roughness-sensitive texture descriptors that encode both frequency energy distribution and directional anisotropy. These handcrafted, physics-informed features reduce data demands and improve interpretability compared to black-box deep learning methods. A hybrid backpropagation neural network (BP NN), designed as a nonlinear function approximator, predicts surface roughness with < 7% error, outperforming conventional statistical methods and run-length matrix-based models. Additionally, a multimodal extension incorporating vibration and acoustic signals demonstrates robustness to machining variations, highlighting the system’s potential for real-time, edge-deployable intelligent quality inspection. This work indicates that integrating physics-driven spectral analysis with machine learning creates a scalable, explainable, and high-accuracy AI solution for Industry 4.0 manufacturing.