Scholarlink

<p>Surface roughness plays a key role in grinding, since this process is usually the last step in machining and directly defines the quality and performance of the final part. Even small changes in surface finish can cause defects or reduce functionality. In this work, we developed a framework to predict surface roughness using industrial edge devices. Data were collected from a Siemens EdgeBox connected to a five-axis grinding machine, recording different signals at 1&#xa0;kHz. After correlation analysis, spindle current was identified as the most important signal. From this, a set of statistical and frequency-based features were extracted. Using 21 measured roughness values, different artificial neural network models were trained and compared. The best model, with a 10–10–10–1 structure, achieved a mean absolute error of 0.095 and prediction accuracy above 95%. Validation was carried out with 100 additional grinding tests, from which 5 new roughness values were measured. The model predictions showed stable performance, confirming the spindle current as a most reliable indicator for surface quality. This study identifies the key signals and features for roughness prediction and prepares the ground for real-time deployment of the system on edge devices in future work.</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

Toward Real-time Prediction of Surface Roughness in Grinding process using Industrial Edge Devices

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月6日

查看全文
  • Seyyed Sajjad Ahmadpoor,
  • Bahman Azarhoushang

摘要

Surface roughness plays a key role in grinding, since this process is usually the last step in machining and directly defines the quality and performance of the final part. Even small changes in surface finish can cause defects or reduce functionality. In this work, we developed a framework to predict surface roughness using industrial edge devices. Data were collected from a Siemens EdgeBox connected to a five-axis grinding machine, recording different signals at 1 kHz. After correlation analysis, spindle current was identified as the most important signal. From this, a set of statistical and frequency-based features were extracted. Using 21 measured roughness values, different artificial neural network models were trained and compared. The best model, with a 10–10–10–1 structure, achieved a mean absolute error of 0.095 and prediction accuracy above 95%. Validation was carried out with 100 additional grinding tests, from which 5 new roughness values were measured. The model predictions showed stable performance, confirming the spindle current as a most reliable indicator for surface quality. This study identifies the key signals and features for roughness prediction and prepares the ground for real-time deployment of the system on edge devices in future work.