Transformer-based network with spatial correlation change and multi-segment attention for sequential EMG recognition.

Abstract

<i>Objective.</i>Surface electromyography (sEMG)-based motion recognition technologies have found widespread application across various fields. While researchers increasingly emphasize the decoding of complex real-world movements, significant challenges remain in sequential feature extraction, application generalization, and other aspects.<i>Approach.</i>To address this, we propose a novel Transformer-based architecture named Multi-Interval Driven Transformer (MIDT). The network introduces a muscle correlation-guided adaptive segmentation module aligning signal partitions with motor phase transitions, and applies hierarchical self-attention mechanisms within and across intervals to capture sub-movement features and long-range dependencies, thereby improving sequential feature modeling. We validated the effectiveness of MIDT on a new sequential movement-based surface electromyography dataset. This dataset includes recording data from 10 subjects and uses 4 types of upper-limb sequential movement paradigms.<i>Main results.</i>The proposed MIDT architecture achieves a classification accuracy of 92.56\% on the ULSE dataset, surpassing state-of-the-art models by 5.07\% with 46.14\% lower cross-subject variance for superior robustness. It also reaches 80.93\% top accuracy on the public dataset, outperforming mainstream methods.<i>Significance.</i>Through visualization analysis, MIDT enables decoding of sub-movement execution and motion state transitions, thereby providing quantitative technical support for personalized and refined motion control and rehabilitation assessment. These findings fully demonstrate the broad application potential of MIDT in wearable human-machine interfaces and neurorehabilitation engineering.