User Intention Estimation for an Elbow Exoskeleton using Surface Myoelectric Pattern Recognition

Abstract This paper presents a neural network for detecting arm movement intentions in an elbow exoskeleton for stroke rehabilitation. The system uses electromyography signals (biceps and triceps), an elbow encoder, and an accelerometer. Data from 23 participants performing six movements resulted in twelve classes (flexion/extension). A CNN processes modality-specific inputs through convolutional and dense layers to extract local and global features, followed by a softmax output. The network distinguishes movements well but struggles with precise target angle estimation. Improvements may require more data or architectural changes, with future work focusing on real-world testing and system integration.