Lower Limb Continuous Motion Recognition Based on DCWT-Residual Attention Network for sEMG Signals

Surface electromyography (sEMG) occurs prior to motion and contains multidimensional information about human body movement, which can be used for classification and recognition of continuous lower limb motions. Traditional feature extraction methods may result in loss of some signal features. To retain as much of the original signal features as possible, we propose a method of continuous wavelet transform combined with convolutional neural networks (DCWT) to fuse information at different scales and capture global features of the signals. For the classification algorithm, we adopt a residual attention network to establish a deep network via residual blocks, capturing more valid features, and adjusting channel weights through attention modules to improve efficiency. Through the continuous motion recognition method of DCWT-Residual Attention Network, we conduct training and testing on classification of five lower limb continuous motions: level-ground walking, walking uphill, ascending stairs, walking downhill and descending stairs. Experimental results show that compared with traditional deep learning algorithms like LSTM and CNN-BiLSTM, our method achieves higher accuracy with a recognition precision of 98.14%.