Brain-controlled, robot-assisted rehabilitation is a promising approach in healthcare, with the potential to significantly enhance and partially automate the recovery of motor systems and the brain structures critical for movement. However, developing an effective rehabilitation system entails numerous challenges and limitations. A primary challenge is the limited data available from individuals recovering from motor function injuries, which is essential for training deep learning models to recognize motor imagery patterns. To address this issue, we present initial experiments using data augmentation and classification results on the collected and augmented dataset. Three different representations of input feature vectors and three augmentation methods were employed. Binary classification involving hand movement and rest, as well as multiclass classification involving left-hand movement, right-hand movement, and rest, were performed. The highest accuracy, 76.00 ± 0.80%, was achieved in binary classification with a CNN classifier without any dataset augmentation, using the time-series representation of the input feature vector.