About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
2026 Technical Division Student Poster Contest
|
| Presentation Title |
SPG-54: Physics-Aware Self-Supervised Learning for CNC Process Monitoring |
| Author(s) |
Shohom Bose-Bandyopadhyay, Francis Ogoke |
| On-Site Speaker (Planned) |
Shohom Bose-Bandyopadhyay |
| Abstract Scope |
As industrial CNC tools become increasingly interconnected and advanced sensing capabilities become more accessible, the volume of generated data makes manual labeling infeasible. Self-supervised learning (SSL) frameworks allow models to learn rich, augmentation-invariant, and information-dense representations without the need for labels, learning to classify based on these representations during fine-tuning. Current self-supervised learning approaches employ generic augmentations that may destroy important semantic meaning during pre-training.
In this work, we isolate the impacts of augmentation selection on the expressivity of the learned representation, and hypothesize that domain-agnostic augmentations (jitter, scaling, permutation) generate less meaningful representations than physics-preserving augmentations (like rotation). We evaluated this using the learned representations to classify temporal data from physics-based time-domain simulations. Our results demonstrate that physics-preserving augmentations consistently outperform domain-agnostic baselines, yielding more robust representations and improved downstream classification accuracy. |
| Proceedings Inclusion? |
Undecided |
| Keywords |
Machine Learning, Surface Modification and Coatings, Modeling and Simulation |