Continuous Estimation of Hand Kinematics from Electromyographic Signals based on Power-and Time-Efficient Transformer...

Surface Electromyographic (sEMG) signals contain motor-related information and therefore can be used for human-machine interaction (HMI). Deep learning plays an important role in extracting motor-related information from sEMG signals. However, most studies prioritize model accuracy without sufficient consideration of model efficiency, including the model size, power consumption, and the computational speed of the model. This leads to impractical power consumption, heat dissipation levels and processing time in wearable computation scenarios. Here, we propose an efficient Transformer method that employs the EMSA (Efficient Multiple Self-Attention) and pruning mechanism to improve efficiency and accuracy concurrently, when estimating finger joint angles from sEMG signals. The proposed method does not only achieve state-of-the-art accuracy but can also be deployed on wearable devices to satisfy real-time applications. We applied the proposed model on the Ninapro DB2-dataset to estimate finger joint angles during grasping tasks. RNN series models, Convolution series models, and Transformer series models were used as reference models for comparison. In addition to common model accuracy, the deployment performance of the models was tested on microprocessors, such as Intel CPU i5, Apple M1, and Raspberry Pi 4B. When tested on 38 subjects of the Ninapro DB2, the proposed model resulted in a correlation coefficient of 0.82 ± 0.04, root mean squared error (RMSE) of 10.77 ± 1.48, and normalized RMSE of 0.11 ± 0.01, which were all similar to the results achieved by the state-of-the-art (SOTA) reference methods. Further, the computational time of the proposed methods was 65.99 ms on the Raspberry Pi 4B, which outperformed all the RNN series models and the Transformer series models. The model size and the power (the minimum size and power are 0.39 MB and 2.28 w) consumption of the proposed model also outperformed that of all reference Transformer methods. These experimental results indicate that our model can maintain the accuracy of the SOTA methods while significantly improving efficiency, thus being a promising approach for real-life applications in wearable devices.