WP-M2 Net: Wavelet-Perception Macro-Micro Dynamic 1D-CNN for Efficient Sequential Motion Recognition from Sparse sEMG.

Abstract

Natural and efficient neuromuscular interfaces serve as a vital bridge connecting next-generation neural engineering with wearable human-machine interaction. While sparse surface electromyography (sEMG) is favored for its portability, its limited spatiotemporal resolution poses significant challenges to recognizing highly dynamic sequential motion tasks such as air-writing. Existing methods typically rely on explicit topological transformations (e.g., pseudo-images and graphs), which are often prone to structural mismatches and struggle to effectively capture the coupled spatiotemporal-frequency dynamics in sparse sEMG. To address this, we reframe sEMG as multivariate time series (MTS) and propose a lightweight, end-to-end dynamic 1D-CNN architecture: the Wavelet-Perception Macro-Micro Network (WP-M2 Net). By bypassing predefined explicit topologies, the network employs 1D convolutions to extract implicit cross-channel interactions, thereby accurately characterizing dynamic muscle synergies. Central to this architecture is the WP-M2 Convolution, a novel dynamic operator inspired by the human "Glance-and-Focus" cognitive mechanism. It decouples feature learning into two synergistic stages: Wavelet Macro-Perception (W-MP) and Micro-Aggregation (MA). Specifically, W-MP integrates the differentiable Wavelet Transform (WT) to efficiently capture global time-frequency priors and generate dynamic weights. These weights subsequently guide MA in content-adaptive local aggregation, effectively balancing macro low-frequency profiles with micro high-frequency details. Extensive evaluations on self-collected and public datasets demonstrate that WP-M2 Net achieves the best accuracy-efficiency tradeoff among mainstream baselines, while maintaining robustness to simulated signal degradation. Moreover, attribution analysis reveals stable cross-subject decision cues aligned with task-relevant muscle activation patterns, supporting the physiological plausibility of the network's recognition basis.