Decomposition of high-density sEMG signals: Extracting multiple spikes from single time windows

Decomposing high-density surface electromyography (HD-sEMG) signals has become a powerful tool in various applications, including prosthetic control and human–machine interaction (HMI). Using deep learning methods to decompose HD-sEMG signals can eliminate preprocessing such as whitening of sEMG, thereby reducing its latency in HMI applications. However, current deep learning methods for blind separation mainly use window-based classification methods, which cannot accurately decompose spike-dense areas. In this paper, we rethink from the perspective of sequence to sequence (seq2seq), define the surface electromyography signal decomposition problem as a regression problem, and propose an HD-sEMG signal decoding method CKC-TCN. This is the first time that the problem of extracting multiple spikes from single time windows has been solved. Rather than class labels, we treat the innervation pulse trains (IPTs) of each motor unit (MU) that are derived from the convolution kernel compensation (CKC) algorithm as continuous time series. We train the temporal convolutional network (TCN) to extract sample-level accuracy IPTs of each MU from the unprocessed HD-sEMG signals, and then extract MU firing time sequences. To evaluate the effectiveness of the proposed method, we conduct experiments on both simulated and real data of HD-sEMG. The results show that, CKC-TCN reduces inference time by 99% compared to traditional methods and improves separation accuracy by over 10% compared to the state-of-the-art deep learning solutions, achieving a significant performance enhancement. This makes it more suitable for real-time applications.