A machine learning-based workflow for automatic detection of anomalies in machine tools

Area of Science:

• Industrial Engineering
• Machine Learning
• Manufacturing Systems

Background:

• Industry 4.0 generates vast sensor data, but practical applications for machine failure detection are limited.
• Existing academic approaches often require extensive training data and simulations, which are impractical for industrial settings.

Purpose of the Study:

• To develop a generalizable end-to-end workflow for detecting production modes and machine degradation states using limited data.
• To address the challenge of insufficient data in industrial machine health monitoring.

Main Methods:

• An integrated workflow encompassing raw data processing, phase segmentation, data resampling, and feature extraction.
• Unsupervised clustering for production mode identification and supervised classification for degradation detection.
• Resampling strategies and classical machine learning models to handle small datasets and identify anomalies.

Main Results:

• The proposed workflow successfully detects anomalies in machine tools using limited data.
• Achieved an average F1-score of nearly 93% in evaluations on a real multi-purpose machine.
• Demonstrated the capability to distinguish between normal and abnormal machine tool behavior.

Conclusions:

• The developed end-to-end workflow offers a practical solution for anomaly detection in industrial settings with small datasets.
• This approach represents a novel contribution, utilizing the entire machine signal for individual tool anomaly identification.
• The findings highlight the potential for improved machine health monitoring and predictive maintenance in Industry 4.0 environments.