Reconstruction of Incomplete Surface Electromyography Based on an Adversarial Autoencoder Network

Surface electromyography (sEMG) signals are often incomplete due to interferences during data measurement, which can degrade sEMG-based applications. To address this issue, this paper proposes a novel adversarial autoencoder model, called the SGMD-AAE, which includes a self-mask generator and a multi-view discriminator. The generator’s binary mask is replaced with a self-mask mechanism, and an adversarial loss is added to promote the reconstruction performance. The multi-view discriminator extracts and fuses deep features of sEMG from time and frequency domains to enhance the generator’s reconstruction ability. The SGMD-AAE model is evaluated on the benchmark NinaPro DB2 database, and the experimental results show that it significantly outperforms incomplete sEMG signals, reducing NRMSE by 88.04% and increasing PSNR by 116.21%. The proposed model also achieves high recognition accuracy for hand gesture recognition even in extreme cases where 90% of the sEMG signals are missing, with an average accuracy exceeding 84%. The effectiveness of the SGMD-AAE model is further verified on a self-collected dataset, demonstrating similar recognition results.