A brain-machine interface (BMI) system allows a person to control robotic devices with brain signals. Among many existing methods for signal acquisition, electroencephalography is the most often used for BCI purposes. Its high temporal resolution, safety to use, portability and low cost are the main reasons for being the most used method. Many situations can affect a person s capability of controlling their body, although brain functions remain healthy. For those people in the extreme case, where there is no motor control, a BCI can be the only way to interact with the external world. Nevertheless, it is still necessary to overcome many obstacles for making the use of BCI systems to become practical, and the most important one is the difficulty to design reliable algorithms for pattern recognition using EEG signals. Inter-subject variability related to the EEG channels and features of the signal are the biggest challenges in the way of making BCI systems a useful technology for restoring function to disabled people. In this paper a method for selecting subject-specific channels and features is proposed and validated with data from seven subjects. Later in the work data is acquired with different EEG equipment and an adapted version of the proposed method is applied aiming online activities. Results showed that the method was efficient for most people and online activities had promising results. It was possible to analyze important aspects concerning the algorithm and inter and intrasubject variability of EEG signals. Also, results showed that it is possible to achieve good results when multiple analyses are performed with the same subject, even with EEG equipment with well known limitations concerning signal quality.