Facial Expression is a very important factor in the social interaction of human beings. And technologies that can automatically interpret and res- pond to stimuli of facial expressions already find a wide variety of applications, from antidepressant drug testing to fatigue analysis of drivers and pilots. In this context, the following work presents a model for Automatic Classification of Facial Expression using as a training base the dataset Challenges in Representation Learning (FER2013), characterized by examples of spontaneous facial expressions in uncontrolled environments. The presented method is composed by a Convolutional Neural Networks Ensemble architecture, using a non-trivial voting system, based on a smart model, Xtreme Gradient Boosting - XGBoost. As performance criteria for validation of the proposed model, were used K-fold and F1 Score Micro techniques to guarantee robustness and reliability of the results, which are competitive with state-of-the-art works.