Machine Learning-Based Measurement Models: A Novel Methodology for GUM-Based Uncertainty Evaluation

The continuous advancement of computational technologies has accelerated the use of data-driven methods across various domains, including metrology. Machine Learning (ML) models, in particular, have demonstrated significant potential to enhance measurement processes, establishing themselves as fully fledged elements of modern metrology. However, in such ML-based measurements, the evaluation of measurement uncertainty requires additional considerations compared with the conventional approaches to measurement. The most widely adopted framework in metrology, i.e., the guide to the expression of uncertainty in measurement (GUM), is based on the assumption that the measurand must be defined in such a way that definitional uncertainty is negligible compared to the other components of measurement uncertainty. In conventional practice, this is achieved by selecting an appropriate relationship between the measurand and a set of input quantities, referred to as the measurement model. In contrast, in ML-based measurements, this relationship is not explicitly chosen by the practitioner but rather inferred from data. Consequently, should adherence to the GUM framework be intended, it is necessary to verify the validity of the GUM assumption and take corrective actions where appropriate. Based on these considerations, this article proposes a novel methodology that enables GUM-based uncertainty evaluation in ML-based measurements. The primary focus is on ML deterministic regression models, whose structure can be readily assimilated into measurement operations. The proposed methodology is applied to two representative case studies. The first case study involves an electric circuit for resistance measurement, where a comparison between ML-based approaches and a consolidated measurement model is provided to ensure consistency. The second case study concerns the estimation of the output power of a power plant, where no conventional measurement model is available. Overall, the proposed methodology enables proper integration of ML practices into the GUM framework, thus broadening the application domain of measurement processes.