An Edge AI Accelerator of LRCN Model with RISC-V Platform for EEG-based Emotion Real-time Detection System

With the development of neural networks and big data, research on emotion recognition has gradually increased. In many ways of emotion recognition, we proposed using Electroencephalography (EEG) signals to achieve high-precision emotion recognition. In this paper, the model we built was based on the concept of Long-term Recurrent Convolution Networks (LRCN), which is used for emotion recognition of EEG signals. In order to realize this real-time wearable system, we employ RISC-V for signal preprocessing and establish the entire system by communicating with our AI accelerator through a communication protocol. In addition, to accelerate and integrate this AI architecture into the RISC-V platform, we optimized the area and computing efficiency of the AI architecture. This optimization improves the data reuse of convolution and fully connected operations and enables acceptance of inputs of different sizes, maximizing hardware reusability. Finally, the AI acceleration chip within the system was implemented on the Kintex-7 platform, achieving an accuracy of 88.6% (two-class classification) and 69.31% (three-class classification) on the SEED dataset and the optimized AI architecture exhibits a power efficiency of 9.26 GOPS/W.