Title:Electromyography Based Cross-Subject Limb Angle Estimation via Hierarchical Spiking Attentional Feature Decomposition Network

Abstract:As human-machine interaction systems are developing towards lightweight and pervasive direction, the role of simultaneous and proportional control (SPC) in human-machine interaction becomes increasingly prominent. However, existing continuous joint angle prediction algorithms based on surface electromyography (sEMG) typically incur high inference costs or are only applicable to specific subjects rather than cross-subject scenarios. Therefore, we proposed a hierarchical Spiking Attentional FEature decomposition Network (SAFE-Net) in order to reduce inference costs and improve recognition accuracy in cross-subject scenarios. This network first encodes the sEMG signals into neural spiking forms through a Spiking Sparse Attention Encoder (SSAE). The compressed features are then decomposed into kinematic features and biological features by a Spiking Attentional Feature Decomposition (SAFD) module. Finally, the kinematic features and biological features are decoded into joint angle values and subject identity, respectively. We validated the effectiveness of SAFE-Net on two datasets (SIAT-DB1 and SIAT-DB2) and compared it with two state-of-the-art methods, Informer and Spikformer. Experimental results demonstrate that, on the one hand, SSAE saves 39.1% and 37.5% power consumption respectively over them in terms of inference costs. On the other hand, SAFE-Net outperforms Informer and Spikformer in recognition accuracy on both datasets. This study showcased that the proposed SAFE-Net can provide accurate predictions in cross-subject scenarios, offering a promising vision for precise continuous control of lower limb rehabilitation exoskeleton robots.