Y-rank can present faults when dealing with non-linear problems. A methodology is proposed to improve the selection of data in situations where y-rank is fragile. The proposed alternative, called k-rank, consists of splitting the data set into clusters using the k-means algorithm, and then apply y-rank to the generated clusters. Models were calibrated and tested with subsets split by y-rank and k-rank. For the Heating Tank case study, in 59% of the simulations, models calibrated with k-rank subsets achieved better results. For the Propylene / Propane Separation Unit case, when dealing with a small number of sample points, the y-rank models had errors almost three times higher than the k-rank models for the test subset, meaning that the fitted model could not deal properly with new unseen data. The proposed methodology was successful in splitting the data, especially in cases with a limited amount of samples.