Efficient AI Model and Hardware Architecture Based on CNN for Arrhythmia Prediction

The development of Edge Artificial intelligence (AI) is increasingly prevalent in real-time recognition, classification, and prediction applications, driven by advances in embedded hardware and lightweight, optimized AI models with low computational requirements. However, deploying lightweight models also faces challenges on hardware with limited resources, including ensuring computational efficiency and utilization, as well as controlling latency and accuracy. In this study, we propose an effective algorithm for the Convolutional Neural Network (CNN) model in arrhythmia prediction applications. Additionally, a hardware architecture is designed to be compatible with this model. The model is optimized through compression and quantization techniques to integer format, making it suitable for edge computing devices. The hardware architecture is implemented using the proposed methods, including the Weight Stationary architecture of the Systolic Array (SA), efficiently organizing data flow, and integrating 8-bit integer computation algorithms for full Processing Element (PE) utilization and resource optimization. Our experimental results can reach over 80% of PE utilization and throughput achieved <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$60.6 \text{GOP}/\mathrm{s}$</tex>.