Centerless grinding is a machining process characterized by highly nonlinear dynamics and large model uncertainty, making it difficult to predict the quality of the worked parts on the basis of the chosen process parameters. Indeed, it is shown that both physics-based and learning-based approaches alone achieve non-satisfactory prediction performance. In this paper a physics-informed learning approach for this problem is presented. It exploits both the prediction of a physics-based (PB) simulation model and a reduced set of experimental data for a data-driven correction. The approach relies on a hierarchical semi-supervised classification, where the training data, classified on the basis of the three quality intervals of interest, are divided in a certain number of sub-clusters w.r.t. the process input parameters (primary features) and enhanced with the classification prediction provided by a physics-based model (apriori knowledge injection). These sub-clusters are then used in the prediction phase, either directly or through a support vector machine predictor. The results on synthetic data provided by a high-fidelity model show an accuracy (Correct Classification Rate) of 97%, vs 94 % of black-box learning methods and 81% of the physics-based model alone.
A Semi-Supervised Physics-Informed Classifier for Centerless Grinding Operations
M. Leonesio;L. Fagiano
2022-01-01
Abstract
Centerless grinding is a machining process characterized by highly nonlinear dynamics and large model uncertainty, making it difficult to predict the quality of the worked parts on the basis of the chosen process parameters. Indeed, it is shown that both physics-based and learning-based approaches alone achieve non-satisfactory prediction performance. In this paper a physics-informed learning approach for this problem is presented. It exploits both the prediction of a physics-based (PB) simulation model and a reduced set of experimental data for a data-driven correction. The approach relies on a hierarchical semi-supervised classification, where the training data, classified on the basis of the three quality intervals of interest, are divided in a certain number of sub-clusters w.r.t. the process input parameters (primary features) and enhanced with the classification prediction provided by a physics-based model (apriori knowledge injection). These sub-clusters are then used in the prediction phase, either directly or through a support vector machine predictor. The results on synthetic data provided by a high-fidelity model show an accuracy (Correct Classification Rate) of 97%, vs 94 % of black-box learning methods and 81% of the physics-based model alone.File | Dimensione | Formato | |
---|---|---|---|
CCTA22_183_rev1.pdf
Accesso riservato
:
Publisher’s version
Dimensione
4.68 MB
Formato
Adobe PDF
|
4.68 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.