Note (General)type:Thesis
The goal of Brain Machine Interface (BMI) systems is to enable humans to interact with computers or machines by using their brain activity. BMI systems capture the user’s brain activity and translate it to a message or a command for certain interactive applications. Systems that make use of BMIs paradigms permit the disabled and elderly people to control wheelchairs, home appliances and robots. Another application is to write sentences and move the cursor on the screen by using brain signals, playing video games, creating arts, etc. Brain Computer Interface (BCI) systems are also investigated on stroke rehabilitation and real time health monitoring.Motor imagery and motor execution are two of the methods that are used to map the EEG signals into external robotic or computer applications. Motor imagery is a dynamic state during which the subjects imagine the performing of an action. In motor execution applications, the subject performs actual movements (for example, moving arms and legs), or other daily life activities. The motor cortex is the part of the brain where the brain signals for limb motions originates. Several studies have compared motor imagery with motor execution. In both paradigms, the brain signals have patterns which can be used for classification.In the recent years, the research literature on BMI systems shifted to implementing Deep Learning (DL) models to map the brain signals into the desired command. In addition, DL has shown good performance in robotics applications. Eliminating one or more of intermediate processing steps, such as preprocessing, feature extraction and classification is one of the advantages of applying DL in BMI/BCI systems. There are two ways to implement CNNs on BMI/BCI systems. In the first method, the raw EEG (electroencephalography) data are directly fed into the CNN. The second implementation is to use the CNN only as a classifier, and employ other algorithms for the feature extraction, such as Short-time Fourier Transform (STFT), or Common Spatial Patterns (CSP) and more advanced versions of it, like Filter Bank Common Spatial Pattern (FBCSP).The deep neural network can do these two functions simultaneously: 1) feature extraction and 2) classification. However, transfer learning has not been yet fully utilized to improve the performance of the deep learning architectures for the BMI systems. Therefore, in this thesis the main motivation is to improve the performance of CNNs by transfer learning. Especially, the impact of transfer learning according to window size and hop size is deeply investigated. Window size is the number of data points taken from the brain signal for classification, while the hop size is the number of data points this window jumps each time. The reason why these parameters are so important is because window size is strongly related to the time it takes to map the user’s brain signals to the robot motion. Therefore, the window size impacts the latency between the command and the robot response. Hop size also impacts the latency due to classification time correlation with the number of augmented samples.We implemented the proposed algorithm in motor and imaginary task using the brain signals. In addition to implementing transfer learning for similar tasks using the same subject, we implemented transfer learning among different subjects. The experimental results show that transfer learning can successfully be utilized to increase the performance of deep learning architectures for brain signal classification applications. A larger window size corresponds to a larger accuracy, but a shorter window size can be utilized by increasing its accuracy through transfer learning. Accuracy can also be increased through a small hop size.The trained CNNs using transfer learning are also implemented to map in real time the brain signals to the robot motion. In the first implementation, the humanoid robot with 18 degrees of freedom developed in our laboratory is controlled using brain signals. We also implemented the imaginary motor task to control the robot action. As stated above there is a tradeoff between window size and robot command accuracy. A small window size leads to faster response of the robotic arm, but the accuracy is not optimal. As the window size gets bigger, the robotic arm responses with higher accuracy, but the motion takes longer to start.This thesis shows that transfer learning can successfully utilized to increase the performance of deep learning architectures on brain signal classification. The results show a satisfying improvement in the CNNs performance.
Collection (particular)国立国会図書館デジタルコレクション > デジタル化資料 > 博士論文
Date Accepted (W3CDTF)2023-03-03T17:22:19+09:00
Data Provider (Database)国立国会図書館 : 国立国会図書館デジタルコレクション