AGMMLN: An Attention-Guided Multiscale Multitask Learning Network for Simultaneous Gesture and Force Level Recognition

Gesture recognition technology based on surface electromyography (sEMG) has attracted widespread attention and research. However, most studies have been limited to gesture recognition at a single force level, ignoring the importance of force level recognition during interaction, leading to limitations in its application. Therefore, an attention-guided multiscale multitask learning network (AGMMLN) is proposed in this article. With the Log-Mel time-frequency spectrum of multichannel nonstationary sEMG signals as input, AGMMLN can recognize gestures and force levels simultaneously and accurately. Given the limitation of fixed-scale analysis of classical convolutional neural networks (CNNs), a multiscale convolution module (MSCM) is proposed to simultaneously extract both fine-grained local features and extensive global features in the temporal spectrum. To filter out redundant information in multiscale features and enhance key features, a group convolution module (GCM) is proposed to facilitate nonlinear interactions between feature channels and enhance feature representation. In addition, to automatically select features beneficial to a specific task from the rich set of shared features, a feature attention module (FAM) is proposed to enhance useful features and suppress useless ones. Experiments are conducted on a self-collected able-bodied dataset and a publicly available amputee dataset to validate the effectiveness of the proposed network. The experimental results show that the proposed method achieves 95.19% and 89.91%, and 90.78% and 75.96% accuracy in the simultaneous recognition of gesture and force levels in able-bodied and amputee subjects, respectively, which is significantly superior to the other compared methods.