This work aims to develop a classifier model based on fuzzy inference rules, which are extracted from support vector machines (SVMs) and optimized by a genetic algorithm. The classifier built aims to diagnose power transformers. The SVMs are learning systems based on statistical learning theory and have provided good generalization performance in real data sets. SVMs, as artificial neural networks (NN), generate a black box model, that is, a model that does not explain the process by which its output is obtained. However, for some applications, the knowledge about how the classification was obtained is as important as the classification itself. Some proposed methods to reduce or eliminate this limitation have already been developed, although they are restricted to the extraction of symbolic rules, i.e. contain functions or ranges in the rules´ antecedents. Nevertheless, the interpretability of symbolic rules is still reduced. In order to increase the interpretability of the rules, the FREx_SVM model was developed. In this model the fuzzy rules are extracted from trained SVMs. The FREx_SVM model can be applied to classification problems with n classes, not being restricted to binary classifications. However, despite the good performance of the FREx_SVM model in extracting linguistic rules, the classification performance of fuzzy classification system obtained is still lower than the SVM, since the partitions (fuzzy sets) of the input variables are predefined at the beginning of the process, and are fixed during the rule extraction process. The goal of this dissertation is, therefore, to extend the FREx_SVM model, so as to enable the automatic adjustment of the membership functions of the input variables through genetic algorithms. To assess the performance of the extended model, case studies were carried out in two databases: iris benchmark and frequency response analysis. The frequency response analysis is a noninvasive and non-destructive technique, because it preserves the characteristics of the equipment. However, the diagnosis is carried out by visual comparison and requires the assistance of an expert. Often, this diagnosis is subjective and inconclusive. The automatic adjustment of the membership functions associated with input variables reduced the error up to 13.38 per cent when compared to the configuration without this optimization. In some cases, the classification performance with membership functions optimization exceeds even those obtained by SVM.