Extensible Cloud-Based Personalized Signal Classification with Small-Scale CNNs: Application to sEMG Gesture Recognition
Year: 2025
Type: preprint
Abstract: This paper presents a cloud-based surface electromyography (sEMG) gesture classification system that leverages personalized, small-scale Convolutional Neural Networks (CNNs), addressing the variability in muscle signals that hinders the adoption of sEMG-based systems. Unlike existing one-size-fits-all sEMG systems, our approach incorporates user-specific sensor and gesture calibration that fine-tunes the model based on individual muscle signal characteristics. This personalization enables the system to adapt to unique signal patterns, making it highly robust to userspecific differences. The system employs a user-dependent number of sEMG sensors connected to a Wi-Fi-enabled device, transmitting real-time muscle activation data to a centralized server for model training and gesture inference. The system creates a personalized model for each new user that allows custom gesture sets. All signal processing, model inference and training happens on the server side in real time. This cloud-based approach reduces the computational complexity required on the device, enabling lightweight and cost-effective hardware. Moreover, the system easily scales to accommodate a large number of users, providing scalability and flexibility.
Authors: Kaavya Tatavarty, Maxwell Johnson, Boris Rubinsky
Institution: University of California, Berkeley
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4407965160