Real-Time Wearable sEMG Onset Detection and Phase Discrimination of Sit-to-Stand Movement via a Compact Dual-Channel DD-CNN

Repeated sit-to-stand and stand-to-sit transitions load the knee extensors and may contribute to work-related musculoskeletal disorders. Reducing this load requires assistive devices and monitoring of knee function, which depend on real-time onset/offset detection and direction-aware classification of each transition. However, no prior wearable surface electromyographic system has delivered this capability for real-time. This study presents a deep learning method that computes both onset/offset detection and direction discrimination of sit-to-stand and stand-to-sit in a developed wearable surface electromyographic system in real-time. Two ESP32-S3 nodes and a hub record from the vastus lateralis and vastus medialis and run a per-burst convolutional detector, while the hub runs a dual-branch classifier with seventeen handcrafted features. Trained offline on the public Gait120 dataset, the networks are deployed unchanged with embedded-firmware parity to the MATLAB reference. Under leave-one-subject-out evaluation on Gait120, the offline classifier separated each transition with 99.6% accuracy and the detector achieved 96.6% completeness. In real-time recordings from thirty healthy adults, the system retained 85.6% classification and 82.0% detection accuracy, with ≈100 ms latency and a 618 KB network footprint. Results show that a low-power wearable delivers combined detection and phase discrimination in real-time, supporting the potential application in assistive-device control and knee-joint monitoring.