This thesis investigates continuous cycling-phase estimation using a single IMU positioned on the thigh, with future application to Functional Electrical Stimulation (FES) control. Two approaches were developed and compared: a geometrical method based on phase-plane analysis and a CNN-LSTM network, first evaluated on synthetic data generated in OpenSim and subsequently on experimental recordings from real participants. The CNN-LSTM showed higher accuracy under several simulated conditions, whereas the classical method provided more consistent performance across subjects on real data. The results highlight the trade-off between interpretability, robustness, and the flexibility of data-driven models, providing a methodological basis for future wearable and adaptive FES systems.
Questa tesi studia la stima continua della fase di pedalata mediante una singola IMU posizionata sulla coscia, con applicazione futura al controllo della stimolazione elettrica funzionale (FES). Sono stati sviluppati e confrontati due approcci: un metodo geometrico basato sul piano delle fasi e una rete CNN-LSTM, valutati inizialmente su dati sintetici generati in OpenSim e successivamente su registrazioni sperimentali di soggetti reali. La CNN-LSTM ha mostrato maggiore accuratezza in diverse condizioni simulate, mentre il metodo classico ha fornito prestazioni più consistenti tra soggetti sui dati reali. I risultati evidenziano il compromesso tra interpretabilità, robustezza e flessibilità dei modelli data-driven, fornendo una base metodologica per futuri sistemi FES indossabili e adattivi.
SINGLE-IMU CYCLING-PHASE ESTIMATION FOR FES CYCLING: GEOMETRIC VERSUS DATA-DRIVEN APPROACHES
GILIBERTI, ANNA
2025/2026
Abstract
This thesis investigates continuous cycling-phase estimation using a single IMU positioned on the thigh, with future application to Functional Electrical Stimulation (FES) control. Two approaches were developed and compared: a geometrical method based on phase-plane analysis and a CNN-LSTM network, first evaluated on synthetic data generated in OpenSim and subsequently on experimental recordings from real participants. The CNN-LSTM showed higher accuracy under several simulated conditions, whereas the classical method provided more consistent performance across subjects on real data. The results highlight the trade-off between interpretability, robustness, and the flexibility of data-driven models, providing a methodological basis for future wearable and adaptive FES systems.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/110636