Prediksi Kekuatan Ujung Jari Berbasis Sinyal EMG Menggunakan Transfer Learning CNN VGG16 untuk Kontrol Genggam Robot Tangan

Abstract

Loss of hand function due to amputation causes significant limitations in daily activities, requiring a prosthetic system capable of producing proportional and adaptive grip control. This study aims to develop a fingertip force prediction system based on electromyography (EMG) signals using a Convolutional Neural Network (CNN) approach with VGG16 and VGG19 architectures, as well as to compare fine tuning and transfer learning strategies for robotic hand grip control. EMG signals were recorded from subjects with healthy hands using MyoWare sensors, then processed through preprocessing stages including filtering, windowing, and transformation to the time-frequency domain using Short-Time Fourier Transform (STFT). The spectrogram representation was converted into a 224×224 pixel RGB image and used as input for the CNN model for the regression task of predicting fingertip force, with actual force data obtained from a force sensitive resistor (FSR) sensor. Model performance was evaluated through offline testing using cross validation, statistical testing using ANOVA, and realtime online testing on a robotic hand system. The test results show that the transfer learning approach tends to produce higher determination coefficient (R²) values in cross validation testing, with the best value reaching 0.90 in certain subjects, but followed by greater prediction errors. Conversely, the fine tuning approach shows more stable performance with lower RMSE and MAE values and better consistency between subjects. In realtime testing, the model was able to directly follow the fingertip force pattern under relatively stable EMG signal conditions, despite a decline in performance due to signal variation and noise. Overall, the results of this study show that the EMG spectrogram-based CNN approach can be used to model the relationship between muscle activity and grip force and integrated into a real-time robotic hand control system. These findings form the basis for the development of more natural, adaptive, and responsive prosthetic hand systems tailored to individual user characteristics.