MyoStep: Feature-Based GNN Model for Estimating Knee Joint Angles by Fusing Signals From sEMG and IMU

Smooth and continuous control over exoskeletons remains challenging for practical applications. Applying surface electromyography (sEMG) and inertial measurement unit (IMU) to predict knee joint angles faces several issues, including complex deployment, precise locating muscles, and equipment interfering with movement. Furthermore, present estimation methods seldom consider the topology of the device. We propose a MyoStep method that applies a feature-based graph neural network (GNN) model to estimate knee joint angles by combining signals from sEMG electrodes and IMU. First, the self-developed leg band collects signals and its sensors are mapped as the graph nodes. Features are extracted and then weighted by the neighborhood component feature selection algorithm, and the top five weighted features are exploited as graph properties. Furthermore, the topological links between the sEMG electrodes and IMU are associated with the edges of the graph, and the mean correlation coefficients between neighboring nodes are computed as the edge attributes. Finally, the graph features are obtained by the ReadOut function of the model and then fed into the fully connected layers to estimate knee joint angles. The sensor deployment of MyoStep is simpler and causes less interference. In addition, the estimated knee joint angles’ root-mean-square error (RMSE), coefficient of determination (R²), and Pearson correlation coefficient (CC) are 2.82°, 0.993, and 0.997, respectively. Compared to the models of convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and GNN-baseline, the R² and CC of MyoStep increase by 18% and 7%, respectively, and the RMSE decreases by 78%. Therefore, the MyoStep method has considerable applicability in simplified deployment and precise control of exoskeleton robots.