Approximative Bayesian approach for uncertainty evaluation in machine learning-based hardness measurement |
|---|
| Junnosuke Takai, Yukimi Tanaka, Masahiro Yoshioka, Katsuhiro Shirono |
- Abstract:
- In recent years, the application of machine learning to the field of metrology has been increasingly explored. In the evaluation of measurement uncertainty using machine learning, it is generally considered necessary to evaluate it through combining two types of uncertainties: aleatoric uncertainty, which arises from randomness, and epistemic uncertainty, which arises from systematic factors. However, the methods for evaluating these uncertainties have not yet been established. In this study, we evaluate measurement uncertainty in the Vickers hardness measurement using a Convolutional Neural Network. For this evaluation, we employ Monte Carlo Batch Normalization as an approximation of a Bayesian Neural Network to evaluate epistemic uncertainty. As a result, it was found that a reasonable evaluation is possible for materials similar to those in the training data.
- Download:
- IMEKO-TC6-2025-004.pdf
- DOI:
- 10.21014/tc6-2025.004
- Event details
- IMEKO TC:
- TC6
- Event name:
- TC6 M4Dconf2025
- Title:
2025 IMEKO TC-6 International Conference on Metrology and Digital Transformation
- Place:
- Benevento, ITALY
- Time:
- 03 September 2025 - 05 September 2025