Convolutional Neural Networks for Raw Signal Classification in CNC Turning Process Monitoring

Area of Science:

• Manufacturing Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Smart manufacturing requires advanced process monitoring for quality control.
• Traditional methods rely on manual feature extraction, which can be time-consuming and less effective.
• Near-real-time prediction of part quality is crucial for optimizing production.

Purpose of the Study:

• To develop a machine learning methodology for near-real-time part quality prediction in CNC turning.
• To enable autonomous feature extraction from raw process signals using a novel deep learning architecture.
• To monitor multiple quality indicators (average roughness, peak-to-valley roughness, diameter deviation) with a single, efficient model.

Main Methods:

• Utilized a one-dimensional convolutional neural network (1D-CNN) architecture for autonomous feature extraction.
• Integrated data from multiple signal channels: vibrations, motor speeds, and motor torques.
• Employed a sliding window technique for continuous monitoring and a classification mechanism to distinguish cutting from non-cutting phases.

Main Results:

• Achieved average F1 scores exceeding 0.97 across all three quality indicators (average roughness, peak-to-valley roughness, diameter deviation).
• Demonstrated robust and consistent performance across various training configurations and design parameters.
• Validated the model's ability to accurately predict part quality using autonomously extracted features.

Conclusions:

• The proposed 1D-CNN methodology offers an efficient and accurate solution for smart manufacturing process monitoring.
• Autonomous feature extraction from raw signals significantly enhances the capability of machine learning models in this domain.
• The developed model provides a compact and effective classifier for near-real-time quality prediction, improving manufacturing efficiency.