An Energy-efficient and Dual-mode AI Accelerator for Wearable Lung Sound Monitoring

Lung diseases seriously impact human health. Using Deep Neural Networks (DNNs) to pre-diagnose lung diseases based on lung sounds is a growing trend. However, most existing DNN hardware accelerators perform direct finegrained multi-category classification for lung sounds, leading to low energy efficiency, as most lung sounds are normal, and only a coarse-grained classification for abnormal detection is sufficient for pre-screening. This paper proposes an energyefficient Artificial Intelligence (AI) accelerator based on a tightly Two-Stage Hybrid Neural Network (TS-HNN) model for wearable lung-sound monitoring applications. First, a TwoStage Data Buffering Scheme (TS-DBS) with a custom Local Memory Bank (LMB) is proposed to support the two-stage acceleration of the TS-HNN model. Second, by observing the channel correlation of two major convolution operators, a Reconfigurable Processing Element Array (RPEA) is designed for efficiently computing standard/pointwise convolution and depthwise convolution, enhancing the computation efficiency and energy efficiency. The FPGA experimental results show that the proposed AI accelerator can achieve an energy efficiency of 96.97 GOPs/W, and it only consumes $0.39 \mathrm{~mJ} /$ frame for normal lung sounds and $2.03 \mathrm{~mJ} /$ frame for abnormal lung sounds, which indicates energy efficiency improvement against the state-of-the-art design by $13.41 \times$ and $2.58 \times$, respectively.