GLCII-DenseNet Integrated With Multiple Blocks for Waist Action Recognition Based on Surface Electromyographic Signals

Abstract

As low-back pain prevalence rises, lumbar rehabilitation robots are becoming more common. By analyzing surface electromyographic (EMG) signals, lumbar movements can be recognized, enabling real-time feedback for personalized training. Deep learning methods are increasingly used to recognize these movements. Conventional deep learning models only capture local spatial information and deep features during the feature extraction process, which limits their ability to extract global temporal dependence features and shallow features. Furthermore, their capability to capture the global contextual information is also relatively restricted. To this end, this article presents a new model that integrates the global–local feature extraction (GLFE) block, the contextual transformer (CoT) block, and a modified backbone architecture of DenseNet: the global and local contextual information integrated denseNet (GLCII-DenseNet) model, designed for the recognition of sparse surface EMG signals from the waist. This model extracts rich shallow, deep, local, and global features, enhancing its capability for feature representation and context feature information capture. It effectively integrates feature information from different locations while extracting a large amount of feature data, thereby improving the accuracy and robustness of EMG signal recognition. To verify the practical validity of the model, we recorded ten healthy subjects, each of whom extracted EMG signals from four muscles while performing six common lumbar exercises. Comparison experiments with other networks show that the model outperforms other methods in recognizing sparse surface EMG signals. In addition, we conducted model generalization comparison experiments to further evaluate the model’s performance. The results show that our model is more robust to noise interference.