FPGA implementation of machine learning algorithms for vibrotactile feedback in prostheses

Abstract

This study aimed to apply discrete event-driven vibrotactile feedback using ma chine learning algorithms in real time. Previously acquired tactile and proprioceptive sensor data were input to an FPGA and classi ed by multinomial logistic regression (MLR) and decision tree (DT) algorithms. Calibrated force and angle values and their derivatives were used as features. Movement-type (stationary, exion, contact, extension, release) and object-type (no object, hard object, soft object) classes were predicted as discrete events. Training of the models was performed in MATLAB o ine; model parameters were implemented in the FPGA by using NI LabVIEW and FPGA module. Vibrotactile feedback stimuli were generated in the FPGA card according to real-time classi cation. FPGA outputs were sent to custom-made power ampli- ers to drive two actuators (Haptuator) placed on both upper arms of participants. The classes were mapped to discrete prosthesis events by using two frequencies and two magnitudes (relative to each participant). Six able-bodied humans participated in psychophysical experiments for measuring absolute detection thresholds and sequen tial pattern recognition of vibrotactile feedback. DT performed better than MLR for both object-type (97% vs. 94%) and movement-type (88% vs. 59%) classi cation in real time. Furthermore, the participants could distinguish vibrotactile feedback signals associated resulting from discrete events with medium recall (0.38 ± 0.08), precision (0.38 ± 0.09), similar to o ine identi cation in our previous work. The presented the sis shows that FPGA implementation of machine learning for vibrotactile feedback is feasible in prostheses. It is expected that human performance for utilizing the feedback may increase during daily use because of additional sensory cues and physical context.|Keywords : FPGA, Somatosensory Feedback, Vibrotactile, Touch, Tactile Sensor, Proprioceptive Sensor, Decision Tree, Multinomial Logistic Regression, Machine Learning, Discrete Event-Driven Sensory Feedback Control.