DESIGN AND IMPLEMENTATION OF AN EEG BRAIN INTERFACE DEVICE UTILIZING SOFTWARE AND HARDWARE COMPONENTS OPTIMIZED FOR EMOTION RECOGNITION TASK

Abstract

This thesis designs and implements a low-cost, real-time EEG-based emotion recognition system using an OpenBCI Cyton board (~$221 total) and eight gold-plated electrodes over the prefrontal and frontotemporal regions. Two deep learning architectures, TSception and the EEG Conformer, were trained on the DEAP benchmark dataset and systematically compared across 16 configurations covering 8 vs. 32 channels and 4–10-second analysis windows. TSception consistently outperformed the Conformer, achieving 66.2% arousal and 57.7% valence binary accuracy at its best setting, with 6–16× stronger correlation with the ground-truth emotion labels. An end-to-end real-time inference pipeline built on Lab Streaming Layer, MNE-Python, and PyTorch delivers arousal–valence predictions with 1–2 seconds of latency. The 8-channel consumer-grade configuration retains most of the 32-channel performance at roughly 1% of the cost of research-grade equipment. Results demonstrate that affordable EEG hardware combined with modern deep learning is a practical path toward real-time emotion-aware applications.