Multi-Physics Monotone Score Transport for Unsupervised Domain Adaptation of Continuous Tool Wear Prediction.

Area of Science:

• Manufacturing Engineering
• Machine Learning
• Materials Science

Background:

• Continuous tool wear prediction across materials is hindered by the need to maintain physical wear scale, not just align sensor features.
• Material shifts in milling distort the relationship between sensor data and tool wear magnitude (flank wear width, VB).

Purpose of the Study:

• To address the challenge of cross-material tool wear prediction by recasting it as monotone wear-scale adaptation.
• To propose and evaluate novel score transport frameworks for accurate continuous tool wear prediction under domain shift.

Main Methods:

• Proposed Multi-Physics Monotone Score Transport (MPMST) framework: constructs a tool-wear score, transports it to the source-domain wear scale, and uses isotonic regression for prediction.
• Evaluated One-Physics Monotone Score Transport (OPMST), a force-only variant of the score transport pipeline.
• Utilized the Mondragon Unibertsitatea-Tool Condition Monitoring (MU-TCM) dataset for two cross-material transfer tasks.

Main Results:

• MPMST reduced mean absolute error by approximately 63% compared to Correlation Alignment (CORAL).
• MPMST achieved approximately 31% lower mean absolute error than a physics-informed Gaussian process baseline.
• Results indicate that the MU-TCM dataset is strongly force-dominated.

Conclusions:

• Monotone score construction and score transport are effective mechanisms for continuous tool wear prediction despite domain shifts.
• The proposed MPMST framework offers a significant improvement over existing methods for cross-material tool wear prediction.
• Force data plays a dominant role in tool wear prediction within the MU-TCM dataset.