Measurement uncertainty propagation in classification machine learning techniques following a Bayesian framework

This paper proposes a framework to propagate and quantify the impact of measurement uncertainties following Bayesian inference on the robustness of machine learning classification accuracy. The use of machine learning techniques in metrology has increased in these last years due to the advancement in sensing technology. However, the data obtained is tainted by uncertainties. The problem emerges when measurement uncertainties may provoke a large region that datapoint could cover, resulting in misclassification. Therefore, the impact of including a priori knowledge (when is available) or Gaussian measurement uncertainties by means of probability class prediction are studied in an industrial case.