TranSEMG: A Trans-Scale Hybrid Model for High-Accurate Hip Joint Moment Prediction

Estimating joint moment using surface electromyography (sEMG) signals is vital in industrial and rehabilitation area. Current prediction methods mostly focused on extracting single-scale temporal features. However, the spatial information between channels is also essential in joint moment estimation. Therefore, we introduced a trans-scale hybrid (TranSEMG) model that has two streams. The spatial-feature-net module was designed for one stream, which can extract the spatial features according to different effects of muscles corresponding to different channels on joint moment. Another stream used the temporal convolutional network (TCN) module to learn temporal features. Then, we utilized the designed fusion module (FM) for temporal and spatial features fusion. The model was tested on experimental data collected from nine healthy subjects walking on treadmill, level ground, and ramp. The root-mean-square error (RMSE), the coefficient of determination (R2), and variance accounted for (VAF) between the predicted and inverse dynamics moment are used to evaluate the prediction accuracy. Compared with the state-of-the-art models, the trans-scale hybrid model can predict joint moments with lower RMSE, higher R2, and higher VAF values. Furthermore, the model also has good versatility and showed good performance in other walking environments, such as ramp and level ground. In conclusion, this study provided a new model to get more features from sequences to predict the hip joint moment and achieved a good prediction accuracy.