An Edge AI System-on-Chip Design with Customized Convolutional-Neural-Network Architecture for Real-time EEG-Based Affective Computing System

In this work, we proposed an edge AI CNN chip design for EEG-based affective Computing system by using TSMC 28nm technology. To improve the performance, Artifact Subspace Reconstruction (ASR) and Short-Time Fourier Transform (STFT) were used for our signal pre-processing and features extraction. The time-frequency EEG feature map was obtained with a multi-channel Differential Asymmetry (DASM) method on 6 EEG channels: FP1, FP2, F3, F4, T7, and T8 according to 10–20 system. The total power consumption of the proposed CNN chip was 71.6mW in training mode and 29.5mW in testing mode. We used 32 subjects data from the DEAP database to validate the proposed design, achieving mean accuracies of 83.7%, 84.5%, and 70.51% for Valence-Arousal binary classification and quaternary classification respectively, showing significant performance improvement over the current related works.