A High Performance Accelerating CNN Inference on FPGA with Arrhythmia Classification

A high-performance artificial intelligence accelerator (AIA) for arrhythmia classification on electrocardiography (ECG) is presented in this paper, which proposes an efficient one-dimensional convolutional neural network (1DCNN) with novel multiplicative behavioral and data reuse. The convolutional layer uses weight stationary (WS) to achieve low memory access on tensor-tensor multiplication (TTM) operations and the fully connected layer uses input stationary (IS) to achieve low memory access on inner product matrix-vector multiplication (IPMVM). The lab database and MIT-BIH arrhythmia database are selected to verify the proposed algorithm. The accuracy of software simulation classification on two databases is 97.3% and 98.3%, respectively. Combined with the hardware implementation of quantization and pruned with the architecture of parallel shift processing element array arrangement (PSPEAA) proposed in this work, the accuracies are 96.6% and 96.5%, respectively. The hardware is implemented on Xilinx PYNQ-Z2, and it takes only 0.233 ms operated at 10 MHz and consumes 0.131 W to classify arrhythmia. Finally, according to the proposed technology, the time of memory access is optimized by 29 times and latency is optimized by 22.5 times compared to using a single multiply-accumulate (MAC). Therefore, the proposed architecture can achieve real-time low-power consumption and high-accuracy arrhythmia classification.