Effects of Image Normalization on CNN-Based EEG–EMG Fusion

Abstract

The rise in popularity of wearable robotic devices has brought many opportunities in the area of assistive devices for rehabilitation. However, despite their numerous advantages, these wearable mechatronic devices are not widely adopted due to a poor interaction between the device and the wearer. To solve this issue, several studies have proposed the use of sensor fusion of electroencephalography (EEG) and electromyography (EMG) signals using Convolutional Neural Networks (CNN) to detect the motion intention of the user and control the wearable device. Although normalization techniques can be applied during data pre-processing in order to improve performance, the correct normalization methods to apply for combined EEG and EMG data are unknown, hindering the potential improvement of the resulting CNN models. Therefore, in this study, no normalization, subject-wise, speed-wise, image-wise, and channel-wise normalization methods were compared. The effectiveness of these methods was tested on the overall and speed specific accuracies of CNN models trained using a database of combined EEG–EMG data collected during elbow flexion–extension motions at different speeds. The results of this study showed that no normalization improved the performance of the overall accuracy of CNN models, with grouped spectrogram-based models achieving an accuracy of 81.57 ± 7.11%. A similar trend was found for speed specific accuracies, in which no normalization showed to be a slightly better option during slow motions. Overall, these results provide information on the different tools that can be used to improve the performance of CNN-based models used for the control of wearable robotic devices.