FPGA-based 1D-CNN accelerator for real-time arrhythmia classification

Zheming Liu, Xiaofeng Ling, Yu Zhu, Nan Wang

Abstract: Nowadays, cardiovascular diseases are prevalent, real-time arrhythmia detection through electrocardiogram (ECG) is a vital aspect of health monitoring. Consequently, there has been growing interest in wearable edge devices capable of real-time ECG classification. Current convolutional neural networks (CNN) for arrhythmia classification often involve a large number of parameters and have high computational complexity. This work introduces a one-dimensional lightweight convolutional neural network model (LW-CNN), which leverages residual connections and one-dimensional depthwise separable convolution (DSC). The proposed network shows accuracy achieving 99.59% on software-implementation with fewer parameters and lower computational complexity. A model compression method combining unstructured pruning and incremental network quantization (INQ) is implemented to further reduce the model complexity. Additionally, a neural network accelerator based on multiplication-free convolutional processing unit is designed with high level synthesis (HLS) to reduce resource consumption and achieve real-time ECG classification. The entire system is implemented on Xilinx Zynq 7Z020 board leveraging PS-PL synergy design and achieves classification accuracy of 96.55%, a latency of 63 ms under 50-MHZ and a power consumption of 1.78W with resource consumption of 13726 LUT, 9 DSP, and 5.5 BRAM, which improves resource efficiency.