Design of a CNN Accelerator for Multitask EEG Signal Classification Based on RISC-V

Electroencephalogram (EEG) has been widely employed in emotion recognition, seizure detection, and brain-computer interface applications. This article presents an elaborated RISC-V-based EEG classification system, targeting on classification tasks, such as emotion recognition and seizure detection. A hardware-friendly convolutional neural network (CNN) model is proposed, which enables automatic extraction of advanced features from EEG signals through combined temporal and spatial convolution operations. A dedicated CNN accelerator is designed to efficiently execute fundamental operations, including convolution, pooling, and fully connected computations. A custom instruction set is developed to coordinate the RISC-V processor with the CNN accelerator for inference tasks. Multiple optimization strategies are implemented, comprising clock gating, dual-voltage domain, and approximate multiplier, to reduce power consumption and area. The system is validated on the FPGA platform. Experimental results demonstrate the peak classification accuracies of 94.98% for four-channel emotion recognition and 97.79% for two-channel seizure detection. Furthermore, the system achieves an improved accuracy of 99.22% when extended to eight-channel emotion recognition. Based on 40-nm CMOS technology and operating at 10-MHz clock, the system chip design consumes estimated 0.572 and 0.328 uJ in emotional and seizure classifications, respectively, with inference latencies of 3.23 and 1.82 ms.