Tackling time-variability in semg-based gesture recognition with on-device incremental learning and temporal convolutional networks

Human-machine interaction is showing promising results for robotic prosthesis control and rehabilitation. In these fields, hand movement recognition via surface electromyographic (sEMG) signals is one of the most promising approaches. However, it still suffers from the issue of sEMG signal's variability over time, which negatively impacts classification robustness. In particular, the non-stationarity of input signals and the surface electrodes' shift can cause up to 30% degradation in gesture recognition accuracy. This work addresses the temporal variability of the sEMG-based gesture recognition by proposing to train a Temporal Convolutional Network (TCN) incrementally over multiple gesture training sessions. Using incremental learning, we re-train our model on stored latent data spanning multiple sessions. We validate our approach on the UniBo-20-Session dataset, which includes 8 hand gestures from 3 subjects. Our incremental learning framework obtains 18.9% higher accuracy compared to a baseline with a standard single training session. Deploying our TCN on a Parallel, Ultra-Low Power (PULP) microcontroller unit (MCU), GAP8, we achieve an inference latency and energy of 12.9 ms and 0.66 mJ, respectively, with a weight memory footprint of 427 kB and a data memory footprint of 0.5-32 MB.