With the advent of industry 4.0, there was a significant increase in the generation of information leveraged by new monitoring technologies. Machine Learning algorithms are widely used in this context for statistical inference, prediction, detection and fault diagnosis. However, redundant data or with low gain of information about the process that is desired to have a prediction or diagnosis, can mean an unnecessary cost, generating slow models. A major challenge, therefore, is to filter this data to capture only a portion that has significant relevance, to optimize resources invested in monitoring systems. This project proposes a Wrapper Feature Selection Method based on a Genetic Algorithm to obtain a subset of the input attributes to provide a satisfactory accuracy in the training and validation of classification models in Industrial Chemical Processes Fault Classification. It was used for Fault Classification and to evaluate the solutions generated by the Genetic Algorithm, a Random Forrest Classifier, from the Ensemble algorithm class. This project used the Tennessee Eastman Process as its object of study. The results were considered promising, with an improvement in accuracy of 1.72% and a reduction of approximately 50% in the number of variables in relation to the base case without selection of variables.