Abstract: Accurate recognition of motor imagery electroencephalogram (EEG) signals is a significant challenge in neuroscience and biomedical engineering. This paper presents an EEG signal recognition system based on a bit-serial convolutional neural network (CNN) accelerator, leveraging its advantages of compact size, low power consumption, and high real-time performance. On the software side, the paper systematically introduces the preprocessing, feature extraction, and classification of EEG data, utilizing Gramian Angular Field (GAF) transformation to map one-dimensional signals into two-dimensional feature maps for network processing. On the hardware side, innovative methods such as column-buffering dataflow and fixed-multiplier bit-serial multiplication are proposed, and a prototype of the bit-serial CNN accelerator is successfully implemented on FPGA. The results show that the FPGA implementation of the bit-serial LeNet-5 accelerator achieves average classification accuracies of 95.68% and 97.32% on the BCI Competition IV datasets 2a and 2b, with kappa values of 0.942 and 0.946, respectively. These performances provide an efficient solution for the recognition of motor imagery EEG signals.