Recognition of Gait Phases from Time-Domain Descriptors of Surface EMG Signals and Dynamic Convolutional Neural
None:
Recent technological advancements facilitate the integration of surface electromyogram (sEMG) signals into control strategies for lower limb prostheses. Particularly, the sEMG signals are used to recognize the motion intent and it is challenging due to the inherent stochastic and nonstationary behavior of the signals. In the present study, a novel approach is proposed for multiclass gait phase classification using a dynamic convolutional neural network (DCNN) of time domain measures of sEMG from hamstrings and quadriceps. For this purpose, sEMG and inertial measurement unit (IMU) data are simultaneously recorded from 20 healthy volunteers walking on a treadmill at a speed of 2.5 kilometres per hour (km/h). The sEMG from four muscles are considered for the analysis namely rectus femoris (RF), vastus lateralis (VL), biceps femoris (BF), and semitendinosus (SEM). Three time-domain descriptors namely root mean square, Wilson amplitude and zero-crossing are used to design DCNN for gait phase classification. The results demonstrate that the proposed framework can effectively distinguish the four gait phases. All the features are found to have significant difference between the four gait phases. The DCNN achieves a maximum accuracy of 94.00%, average sensitivity of 94.25%, average specificity of 97.75% and Mathews correlation coefffient (MCC) of 92.00%. The findings indicate that the proposed approach effectively decodes the motion intent of lower limb muscles, potentially paving the way for the development of more precise movement control in lower limb prosthetics.
