Abstract Entropy has been positioned as a measurement capable of detecting irregularities and uncertainties in time series. Specifically, Multiscale Permutation Entropy (MPE) has positioned itself as an alternative to classical characterization methodologies, managing to identify nonlinear dynamics present in non-stationary signals. MPE has been used successfully to identify states of different systems without relying heavily on factors such as length or sampling frequency. This article presents a robust methodology that combines MPE with a feature selection technique and a supervised classifier (Analysis from Variance (VRA) and K-Nearest Neighbours (KNN) respectively). The robust proposed methodology is validated through two experiments (bearing failures and states of an internal combustion engine (ICE)). The proposed methodology overcomes the experiments with high precision and low computational consumption.