Confidence Interval Assessment for Charpy Impact Energy Predictions – A Gaussian Mixture Model (GMM)-Based Approach

Abstract One of the obstacles of applying data-driven models in industry is the lack of confidence in the accuracy of the model predictions. One way of overcoming this is by adding a confidence interval on the predictions. In this paper, we propose a new approach for the confidence interval assessment for model predictions based on a Gaussian Mixture Model (GMM) framework. The advantages of the presented approach include its capability to handle complicated non-white noise sequences, the ability to provide an accurate confidence interval, and its independence to the type of the data model to be assessed. The proposed approach is applied to an industrial case study: the Charpy impact energy prediction, which includes real industrial data containing a significant noise component and with an inherited sparse data distribution. The resulting confidence intervals reflect the prediction uncertainty against the test data. Furthermore, the GMM-based approach can also be used for model bias correction. The GMM-based confidence interval assessment for data driven models represents a valuable contribution especially in the case of critical applications.