A Brain Computer Interface (BCI), is a system created for performing communication and control of computational devices by the analysis of brain signals. Particularly, Motor Imagery (MI) based BCI systems allow a real-time interaction by using the electrical signals that are generated by the brain when the user imagining certain movements or actions. Acquisition of these signals can be done by invasive and non-invasive devices, among which, electroencephalography is a non-invasive technique that is widely used. It is based on the super cial placement of the electrodes on the scalp, which avoid afecting the health and well-being of potential users, also is a portable technology and lower-cost than other alternatives. From the diferent stages required to develop a BCI, the signal characterization and the classi cation tasks continue to be the main research challenges, since the performance of the whole system depends on them. This thesis proposes the use of convolutional neural networks (CNN) for the classi cation of electroencephalographic signals, in order to identify the action imagined by a person. The proposed network architecture is trained with representations of the spectral power density (PSD) of the signals; and the hyperparameters of the network are de ned by a metaheuristic optimization algorithm, which obtains the best accuracy in the signal classi cation in training stage. The proposed approach was e valuated using two public and well-known databases, i.e. BCI Competition IV 2a and BCI Competition IIIa. According to the results, this approach provides a reliable strategy to diferentiate movement imagination, outperforming state of the art that used the same datasets. These results demonstrate that this is a valuable and promising strategy for the design of brain computer interfaces