부산대학교 도서관

EEG-based Brain-machine Interface (BMI) has been intensively investigated for the past few years. This kind of system allows the humans who lost their mobility to interact with the surrounding environment. Robust BMI systems are achieved by utilizing Deep Learning such as CNN, LSTM etc. to recognize the user intention from EEG signals. Nonetheless, the BMI system still poses a crucial challenge in achieving robust systems. First, EEG signals are heavily contaminated with the noises caused by muscle movement, eye movements, etc. Not to mention the subject orientation of each subject when performing the same tasks. This may elicit the feature of the target signals. Second, from the Deep learning itself. Deep learning requires a huge amount of data for the model to be trained on. Together with high variation signals such as EEG, achieving robustness proves to be a difficult task. To target these issues, we combine transfer learning in CNNs and Genetic algorithms to improve the BMI classification rates and improve the training data and time. We performed the experiments to 1) Optimize the channels to reduce information redundancy and 2) Select the best subject's data to train the base model before transferring the pre-trained model to train the target subject. The results show an improvement over non-optimization algorithms for both experiments.