Physics-informed Gaussian process for tool wear prediction.

Area of Science:

• Mechanical Engineering
• Machine Learning
• Manufacturing Technology

Background:

• Tool wear monitoring (TWM) is crucial for CNC machining quality and safety.
• Data-driven TWM requires extensive data and complex models.
• Physics-based TWM models struggle with adaptability to varying conditions.

Purpose of the Study:

• To develop a novel physics-informed Gaussian process model for accurate tool wear prediction.
• To integrate the strengths of physics-based and data-driven TWM approaches.
• To enable effective TWM with limited available sensor data.

Main Methods:

• Developed a physics-informed Gaussian process regression (PB-GPR) model incorporating three physical tool wear models.
• Constrained the Gaussian process mean function using wear models for improved physical relevance.
• Utilized multi-sensor signals and multi-domain features for model training and updates.
• Enabled small-data training and model adaptation with new measurements.

Main Results:

• The PB-GPR model demonstrated significant improvements in tool wear prediction accuracy.
• The proposed approach showed enhanced robustness in extrapolation compared to conventional methods.
• Validation through high-speed milling experiments confirmed the model's effectiveness.

Conclusions:

• The physics-informed Gaussian process model offers a superior approach to TWM in CNC machining.
• This method effectively balances physical principles with data-driven learning for practical applications.
• The PB-GPR model provides a robust and accurate solution for real-world tool wear prediction challenges.