Generative deep learning for lower limb motion prediction with a small training EMG data set

• Main Author: Huang, Zhanfeng
• Other Authors: Lin Zhiping
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• Online Access: https://hdl.handle.net/10356/168371
• Institution: Nanyang Technological University

Human-Robot interaction rehabilitation systems have attracted widespread atten- tion among researchers and medical practitioners because they can help disabled people regain their mobility. Surface electromyography (sEMG) signals are ag- gregated muscle action potentials measured on the surface of the skin. Due to diversified EMG activation patterns within and across individual subjects, as well as a large number of different actions, it is often prohibitively expensive to collect massive data sets from each individual that cover each possible ac- tion class in real-world use scenarios. This dissertation uses deep convolutional generative adversarial networks (DCGANs), to design a method to learn and generate similar sEMG signals to a particular class, while the model can learn the concept of this class with others. We also analyzed the performance of this method in motion prediction and found that the discriminator in the model can identify EMG signals kicking in different directions with an accuracy of 89.31% ± 6.52. Finally, we propose to use dynamic time warping (DTW) and fast Fourier transform mean square error (FFT MSE) to evaluate the signal quality from the time domain and frequency domain respectively, and find that the signals generated by the model have a similarity of more than 95% in these two indicators. The research in this paper explores a data augmentation method applied to a small EMG dataset, which provides a broad direction for the sub- sequent research of motion-assisted robotic systems.