FEATURE EXTRACTION BASED ON SIMILARITY MEASUREMENT IN ELECTROENCEPHALOGRAPH SIGNALS OF BRAIN COMPUTER INTERFACE

• Main Author: Rio Hermanto (NIM 33212005), Beni
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/26071
• Institution: Institut Teknologi Bandung

<p align="justify">The nervous system is a part of the system of human organs, which has three (3) main functions, namely: receives sensory input by gathering information from the sensor receiver, the process of integration process and interpret the input received by the sensor, and produce output motor to perform activation affect the organs in terms of providing a response. In the nervous system, the brain is one organ where nerve cells are part of the brain with a smaller scale. The number of nerve cells in the brain approximately one hundred (100) billion cells. The activities of the nerve cells in the scale of the rest of the ions generate a potential difference of about -70 millivolts. Overall brain activity associated with nerve cells that produce a cumulative potential difference that can be detected on the outside of the head in the range of 20-100 microvolt. The device used to measure any potential difference in the outer part of the head known as Electroencephalograph abbreviated EEG. <br /> <br /> <br /> The main function of the nervous system which is owned, there are several types of disorders, one of which can affect the nervous system paralysis which could not continue to command motion or motor function. This paralysis can affect the ability of human beings in terms of communication or mobility. To be able to help <br /> <br /> people in a state of paralysis that needed a tool, the system of Brain Computer Interface (BCI). BCI system captures brain activity to be translated directly into a <br /> <br /> message or command. To capture brain activity one can utilize EEG device, so that the EEG-based BCI system is a translation of brain signals into a form of a message or command. At BCI system there are two (2) of the main processes: feature extraction and classification. EEG-based BCI systems with a focus on signals associated with movement, feature extraction process is choosing the features of brain signals to be used as a parameter to represent several different movements; and the classification process is to determine whether the feature is included in the list of classification of the movement. <br /> <br /> <br /> The determination of the signal feature to be used as an indicator of motion intentions can be taken on the time domain signal or frequency domain. The signal features in the time domain can be obtained using the cross-correlation approach, which is the product of the reference signal on a particular EEG channel and other channel signals. Reference signals can be selected directly from the available EEG channels. The cross-correlation approach produces a value series called cross-correlation sequence. Feature extraction can use 6 <br /> <br /> statistical basic values (maximum, minimum, average, median, mode, and standard deviation) of cross-correlation sequences, for subsequent use as data dimensions in the classification process. <br /> <br /> <br /> In this study, the initial stage of the process performed on the EEG signal is the application of bandpass filter at 8-35 Hz. The frequency range used in the filter taking into account the typical signals associated with the motion intentions, namely mu and beta rhythm. The determination of the reference signal used in the cross-correlation method is dynamic so that each EEG channel is treated as a reference signal alternately to obtain the most optimal channel in obtaining good accuracy. In addition to reference signals, on features that can use 6 statistical basic values, independent testing of each feature is performed. Feature testing is performed to be able to analyze whether the feature contributes to the achievement of good accuracy or not. Trial data used, namely BCI Competition III, IVa dataset with 10-fold cross validation as validation. <br /> <br /> <br /> <br /> The results obtained, that the initial process with the application of bandpass filter on 8-35 Hz to contribute increased accuracy for the better. The reference signal used to produce a high degree of accuracy is dynamic, where the EEG channel as a reference signal depends on the subject. This proves that the reference signal cannot be determined by selecting directly one of the EEG channels but required testing to obtain the optimal channel for achieving high accuracy. The most optimal determination of EEG channels as reference signals in the calibration process can help simplify channel usage on the real usage side. In the study of dimensional determination of the feature, also found that to obtain <br /> <br /> good accuracy is not required many dimensions but simply use the maximum value of the cross-correlation sequence. The maximum value is a reflection of similarity of signal that can be used as a single feature in the classification process. <br /> <br /> <br /> The accuracy of the approach used in data BCI Competition III, IVa dataset obtained an average accuracy of 99%. For other data from ten subjects obtained an average accuracy of 84%. With accuracy results obtained in the experiment in two brain signal data, feature extraction approach based on cross-correlation maximum value as a single value and the selection of appropriate reference signal is a reliable approach to be used as a base in the feature extraction BCI system as a whole.<p align="justify">