Action Intention Recognition based on L2 regularized LSTM-CNN
Jiazheng Hu, Qi Wang · 2024
With the advancement of industrial technology, intelligent manufacturing, and Industry 4.0, industrial exoskeleton technology has emerged as a pivotal component within the realm of intelligent manufacturing. Electromyographic signals (EMG) play an indispensable role in this domain. Given that the torso serves as a vital hub for human activities, accurate intention recognition holds immense significance in augmenting industrial production and facilitating the application of exoskeleton technology. In this study, we propose a hybrid algorithm framework that combines Long Short-Term Memory network (LSTM) and Convolutional Neural Network (CNN) for intent recognition of human body behavior. By collecting EMG data generated by healthy adult volunteers during specific tasks and normalizing them using fast Fourier transform, CNN is employed to extract spatial features of EMG signals while LSTM captures their temporal dynamics. Furthermore, L2 regularization is incorporated to effectively mitigate overfitting. Experimental results demonstrate an average accuracy rate of 97% in intent recognition tasks, showing excellent performance and potential practical applications in intelligent manufacturing and industrial exoskeleton.