AI Summary & Key Takeaways
AI InsightResults show that accurate force estimation is possible without anatomically precise sensor placement, reducing hardware complexity and cost while keeping performance suitable for gait analysis and wearable health applications.
📌 3 Key Takeaways
Abstract
This paper presents a smart insole that combines flexible TPU optical fiber sensors with force-sensitive resistors (FSRs) on a 3D-printed TPU base to estimate plantar forces during walking. Three thermoplastic polyurethane optical fibers, illuminated by red lasers and read by light-dependent resistors, were routed in a non-anatomical, irregular (‘random’) layout and compared against six FSR channels taken as reference targets. Signals were sampled and streamed via an ESP32 microcontroller over Bluetooth. Using a sliding-window approach (20 samples), simple statistical features from the three optical channels were used to train supervised regressors—Gradient Boosting, Adaptive Boosting, and a shallow artificial neural network—to predict each FSR output. Across sensors, models achieved R² between 0.865 and 0.951 and mean absolute error (MAE) between 29.0 and 48.9. Adaptive Boosting gave the lowest average MAE and stable R², while the artificial neural network reached the highest R² for several regions. Results show that accurate force estimation is possible without anatomically precise sensor placement, reducing hardware complexity and cost while keeping performance suitable for gait analysis and wearable health applications.
Özet
This paper presents a smart insole that combines flexible TPU optical fiber sensors with force-sensitive resistors (FSRs) on a 3D-printed TPU base to estimate plantar forces during walking. Three thermoplastic polyurethane optical fibers, illuminated by red lasers and read by light-dependent resistors, were routed in a non-anatomical, irregular (‘random’) layout and compared against six FSR channels taken as reference targets. Signals were sampled and streamed via an ESP32 microcontroller over Bluetooth. Using a sliding-window approach (20 samples), simple statistical features from the three optical channels were used to train supervised regressors—Gradient Boosting, Adaptive Boosting, and a shallow artificial neural network—to predict each FSR output. Across sensors, models achieved R² between 0.865 and 0.951 and mean absolute error (MAE) between 29.0 and 48.9. Adaptive Boosting gave the lowest average MAE and stable R², while the artificial neural network reached the highest R² for several regions. Results show that accurate force estimation is possible without anatomically precise sensor placement, reducing hardware complexity and cost while keeping performance suitable for gait analysis and wearable health applications.