Scholarlink

This paper presents an approach to automatic anomaly detection in CNC mill data combining machine learning and neural network classification methods. The input material was an open dataset of CNC Mill Tool Wear containing 25286 observations of vibration and process parameters collected during 18 experiments. In the preprocessing stage, the data are normalized and reduced by the principal component method for visual analysis. Next, the HDBSCAN density algorithm marks points that deviate significantly from the cluster centers as potential anomalies. The resulting labeling is used to train a multilayer perseptron with ReLU activation functions and sigmoid output. The model achieved 93.4% accuracy, 97% precision and 93% F1-score on a test subsample, which is comparable to the results of the LSTM model from recent studies at significantly lower computational cost. The presented pipeline is easily adaptable to other data sources and can serve as a basis for predictive maintenance systems for mechanical engineering equipment.

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

Software Module for Anomaly Detection in CNC Milling Using Clustering-Based Machine Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • R. A. Nezhmetdinov,
  • I. A. Kovalev,
  • V. V. Ermakov,
  • N. Y. Chervonnova,
  • D. V. Nikitin

摘要

This paper presents an approach to automatic anomaly detection in CNC mill data combining machine learning and neural network classification methods. The input material was an open dataset of CNC Mill Tool Wear containing 25286 observations of vibration and process parameters collected during 18 experiments. In the preprocessing stage, the data are normalized and reduced by the principal component method for visual analysis. Next, the HDBSCAN density algorithm marks points that deviate significantly from the cluster centers as potential anomalies. The resulting labeling is used to train a multilayer perseptron with ReLU activation functions and sigmoid output. The model achieved 93.4% accuracy, 97% precision and 93% F1-score on a test subsample, which is comparable to the results of the LSTM model from recent studies at significantly lower computational cost. The presented pipeline is easily adaptable to other data sources and can serve as a basis for predictive maintenance systems for mechanical engineering equipment.