Comparison of Finger Motion Classifiers With Respect to Number and Location of Tactile Sensors

This study investigates the performance of AI classifiers in classifying motions of individual fingers for prosthetic hand using a varying number of tactile sensors. Specifically, we compare the classification rate of Convolutional Neural Network (CNN) and Linear Support Vector Machine (LSVM) with Polyvinylidene Fluoride (PVDF) tactile sensors placed on key forearm muscles. Our study reveals that CNN achieves significantly higher classification rates than LSVM when using fewer sensors, with a notable improvement in accuracy from 65.8 % (LSVM) to 72.8 % (CNN) with one sensor. However, as the number of sensors increases to three or four, the performance difference between CNN and LSVM diminishes, both reaching approximately 95 % accuracy. This suggests that while CNN are advantageous for scenarios with fewer sensors due to their superior noise tolerance and feature extraction capabilities, the choice of classifier becomes less critical with a larger sensor array. These insights can guide the design of high degree of freedom motorized prosthetic hand by minimizing the sensor number while maintaining high classification accuracy.