Scholarlink

<p>Tool condition monitoring is essential for improving machining quality and reducing production costs. In recent years, numerous tool condition monitoring methods have been developed, but most employ supervised learning. However, supervised learning requires large amounts of labeled training data, necessitating extensive machining experiments. To address this challenge, we propose a tool wear prediction method combining an unsupervised learning approach (One-Class SVM) with a tool wear prediction formula incorporating cutting speed, feed rate, depth of cut, and minimum quantity lubrication (MQL) condition indicators. This method enables tool wear prediction using only a small amount of labeled data. This study targets conditions where multiple different workpiece materials are machined consecutively using the same tool. We focus on the degradation of prediction accuracy caused by variations in cutting force, vibration, and acoustic signal features due to material changes and lubrication conditions. This study addresses the variation in feature distribution accompanying material changes by introducing feature selection using Spearman’s rank correlation coefficient, anomaly correction considering cutting parameters and MQL conditions, and a method to redefine normal data for each workpiece material. As a result, we confirmed a significant improvement in tool wear prediction accuracy even after material changes. Furthermore, quantitative evaluation using RMSE showed that the proposed method achieved an RMSE of 16.9&#xa0;μm, demonstrating prediction accuracy comparable to the supervised learning model’s (Light GBM) 16.4&#xa0;μm. This demonstrates that tool wear prediction based on unsupervised learning using only a small number of training data is practically effective even in real machining environments.</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

End mill tool wear prediction method using a small amount of labeled data and one-class support vector machine

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Takahiro Yoshizumi,
  • Hiroshi Murakami,
  • Akio Katsuki

摘要

Tool condition monitoring is essential for improving machining quality and reducing production costs. In recent years, numerous tool condition monitoring methods have been developed, but most employ supervised learning. However, supervised learning requires large amounts of labeled training data, necessitating extensive machining experiments. To address this challenge, we propose a tool wear prediction method combining an unsupervised learning approach (One-Class SVM) with a tool wear prediction formula incorporating cutting speed, feed rate, depth of cut, and minimum quantity lubrication (MQL) condition indicators. This method enables tool wear prediction using only a small amount of labeled data. This study targets conditions where multiple different workpiece materials are machined consecutively using the same tool. We focus on the degradation of prediction accuracy caused by variations in cutting force, vibration, and acoustic signal features due to material changes and lubrication conditions. This study addresses the variation in feature distribution accompanying material changes by introducing feature selection using Spearman’s rank correlation coefficient, anomaly correction considering cutting parameters and MQL conditions, and a method to redefine normal data for each workpiece material. As a result, we confirmed a significant improvement in tool wear prediction accuracy even after material changes. Furthermore, quantitative evaluation using RMSE showed that the proposed method achieved an RMSE of 16.9 μm, demonstrating prediction accuracy comparable to the supervised learning model’s (Light GBM) 16.4 μm. This demonstrates that tool wear prediction based on unsupervised learning using only a small number of training data is practically effective even in real machining environments.