When considering data-driven modelling, uncertainties, errors and inconsistencies in the data can more often than not lead to sub-optimal predictions. A new framework based on rough sets theory is proposed and applied to an industrial data set obtained from a Charpy impact energy test for alloy steels. The inconsistent/consistent data sets are then used to train a series of artificial neural networks (ANN) for Charpy impact energy prediction for alloy steels. A k-nearest neighbor is used to classify the data points; if an object is classified as consistent, the ANN trained with the consistent data set provides a single prediction while if the object is classified as inconsistent, several ANN trained with different sets of inconsistent data are used to provide an interval prediction. Experimental results show an improvement in the consistent data set compared with a benchmark model. Also, the interval prediction provided by the various ANNs in the inconsistent data set represents a better alternative to the single point prediction results.
