Brain Computer Interface (BCI) is defined as machines using the brain signals for its control. The motor neural activity of the brain can be recorded from the human scalp using EEG recording equipment and converted into control commands representing the needs of the person. These can be used for the control of devices such as a prosthetic arm, a joystick or a wheelchair and may be very useful for persons with physical disabilities. The performance of such system, however, depends heavily on the quality of the recorded signals and the subsequent features extracted from it. This research proposes a new protocol for experimental setup, frequency band as well as channel selection approach based on raw signal energy to improve data dimension reduction process. A new protocol for acquiring brain EEG signal using 19 non-invasive scalp electrodes were developed. The frequency bands related to the motor actions, namely alpha1 (8-10 Hz), alpha 2 (11-12 Hz), beta 1 (13-15 Hz), beta 2 (16-18 Hz) and beta 3 (19-25 Hz) were extracted using a customized filter. A novel four class brain computer interface (BCI) based on four tasks motor imagery signals was also designed. Classifications of the tasks using three spectral and fractal-based features with three different types of neural networks were performed. A novel method to minimize the number of frequency bands and EEG channels for the classification of tasks is proposed without sacrificing the classification accuracy. The performance of the neural network models with features from the selected channels were compared to that with all the 19 channels. The proposed approach with limited number of chosen channels were able to achieve a minimum average accuracy of 90% similar to that of the present methods but with reduced computational times. The proposed method also has the advantage of being able to identify the most discriminating regions for the EEG electrode channels for motor imagery tasks classification. The results show that the proposed frequency band selection and channel reduction method is a viable data dimension approach with reduced computational time to classify a motor imagery task for a motor imagery BCI system. This research work will reduce the computational load, and hence, time in BCI systems. This opens up the possibilities for the migration of pc-based to microcontroller-based implementations for better mobility.
