Methodological challenges of outlier detection in metrology using machine learning models

The paper addresses the challenges associated with applying machine learning models to detect outliers in metrological datasets. While such models ensure the possibility to identify complex deviations in the structure of a sample without relying on prior statistical assumptions, they do not provide normatively justified criteria for assessing the reliability of their decisions. Specifically, such models lack interpretable confidence indicators, metrological traceability, and formalised thresholds to determine whether an outlier is genuine. One proposed solution involves assessing the impact of eliminated anomalous values detected by the Isolation Forest model on the standard measurement uncertainty of Type A when the initial sample size is preserved through repeated measurements. This approach was validated using real-life measurements of liquid flow performed with Coriolis flowmeters of various diameters. The results empirically proved the effectiveness of the method in cases where the elimination of distortion-inducing values led to a significant reduction in measurement variability. However, several limitations were also identified, including the sensitivity of models to small sample sizes, the impracticality of performing repeated measurements in many real-life scenarios, and the lack of an objective threshold to determine the “significance” of uncertainty reduction. These findings highlight the need for further study of the formalization of confidence criteria in anomaly detection within the metrological domain, particularly in the context of compliance with international standards such as ISO/IEC 17025. Despite these limitations, the application of machine learning models opens new opportunities for automating the analysis of metrological data and highlights the need to develop harmonized approaches for integrating such solutions into the regulatory framework.