LLANet: An Automatic Framework for Efficient Prediction of Lower Limb Activities Using sEMG Signals

Wearable technology has increased the usage of surface electromyography (sEMG) signals in recent years. Recognizing lower limb activity more accurately is necessary for developing mechanical lower limbs or exoskeletons controlled by the nervous system. EMG signals detect the electrical activity of the muscles involved in human activities. In this letter, we proposed LLANet for lower limb activity prediction using an sEMG signal. The proposed work is on a publicly available UC Irvine dataset, i.e., EMG Physical Action Dataset. Conventional machine learning techniques necessitate multiple manual procedures, including decomposition, feature extraction, and classification. To address these constraints, developing a framework using a short-time Fourier transform (STFT) based on synchrosqueezing coupled with convolutional neural networks (CNN) called lower limb activity prediction CNN (LLANet) is proposed. First, the EMG signal is converted to spectrogram images using an STFT, followed by the synchrosqueezing transform to enhance the time–frequency representation. Extracted spectrogram images are fed to CNN models. Subsequently, we explored deep learning models, including CNN, long short-term memory, and STFT+CNN. In addition, the experimental results show that deep feature extraction using LLANet achieved the highest accuracy of 98.56% compared with other models. We also compare the propose model with three time–frequency image, such as continuous wavelet transform (CWT), STFT, Stockwell transform, and log-Mel spectrogram.