Lower-limb motion recognition using Gramian angular difference field and multi-scale attention ResNet18 for sEMG signal

Abstract

Aiming at the problem that traditional methods are insufficient in capturing temporal dependencies for surface electromyography signals, an integrated lower-limb motion recognition framework combining Gramian angular difference field and multi-scale attention ResNet18 is constructed in this paper. First, the Gramian angular difference field is introduced to encode one-dimensional surface electromyography signals into two-dimensional images, which preserves the temporal order and amplitude information of the original signal. Second, a multi-scale attention ResNet18 is designed, integrating a multi-scale inception block, a convolutional block attention module, and an adaptive multi-scale feature fusion module. It captures the spatiotemporal features of different granularities and enhances discriminative feature responses. Experimental results show that the proposed method achieves an accuracy of 93.32%, representing a 2.26% improvement over the baseline ResNet18. Ablation experiments confirm the synergistic effect of each module, and visual analysis reveals the intrinsic relationship between feature representation and human motion physiology. These results demonstrate the feasibility of the proposed framework as a proof-of-concept under the current intra-subject evaluation, while cross-subject and temporal validation remain essential future work.