The calibration of gas flow meters at the Italian National Institute of Metrological Research (INRIM) is constrained by a manual and subjective validation workflow, leading to significant operational inefficiencies and potential inconsistencies in ensuring metrological traceability. This thesis addresses this challenge by conducting a comparative study of data-driven anomaly detection models, evaluating robust statistical, unsupervised, and supervised machine learning approaches on historical calibration data from INRIM's BellGas primary standard. Building on this analysis, the research presents the development and integration of the optimal detection model into a software tool within the existing LabView environment, enabling real-time, automated feedback for operators. The evaluation, based on robust metrics like the F1-Score, demonstrates the superiority of supervised models and reveals that feature engineering, specifically temporal differencing, is the most critical factor for achieving high performance.