Most of the approaches of feature extraction for data-driven rotating machinery fault diagnosis assume characteristics of periodicity and seasonality typically inherent to linear signals obtained from different sensors. Nevertheless, the behavior of rotating machinery is not necessarily linear when a failure occurs. Thus, new techniques based on the theory of chaos and nonlinear systems are needed to extract proper features of signals. This article introduces the use of features extracted from the Poincaré plot (PP), which are computed over vibration and current signals measured on a gearbox powered by an induction motor. A comparison between the performance of classic statistical features and PP features is developed by applying feature analysis based on analysis of varaince (ANOVA) and cluster validity assessment to rank and select the subset of best features. K-nearest-neighbor (KNN) algorithm is used to test the performance of the selected feature set for fault severity classification. The use of PP for the analysis of nonlinear, nonperiodic signals is not new; however, its application in mechanical systems is not widely extended. Our contribution aims at highlighting the use of the PP features, supported by data collected from a test bed under real conditions of speed and load, to proof the potential application of this approach. The results show that PP features extracted from the current signal yields 96% of classification accuracy when using at least 11 features.