Interpretable analysis of feature importance and implicit correlation based on sEMG grayscale. images

For patients requiring upper limb rehabilitation, the hand rehabilitation robot assists the patient in completing movements within a certain training trajectory to achieve therapeutic results. There have been studies based on deep learning to convert surface electromyography (sEMG) signals into sEMG images for motion intention analysis. Although good recognition accuracy has been achieved, the working principle of neural networks and the processing of image features by the networks are not well explained. The interpretability of deep neural networks determines human confidence in neural network decisions. In this paper, we design a method based on feature importance and implicit correlation for hand motion intention recognition, experimentally explored that convolutional neural networks have implicit definitions for sEMG grayscale images of the same hand gesture action, and verified the effectiveness of the designed method.