Zynq-Based Deep Learning for Lightweight Recognition of Acetone Concentration in FAIMS System

In the quantitative analysis of acetone concentration using high-field asymmetric ion mobility spectrometry (FAIMS), the FAIMS spectral dataset exhibits low variability, and the number of available samples is limited. In this article, we propose that a low-complexity deep learning algorithm is efficiently implemented on a Zynq platform, which integrates an advanced RISC machine (ARM) processor and field-programmable gate array (FPGA), using a hardware and software co-design approach. To evaluate the feasibility of the hardware implementation of the model, we compare the number of parameters and recognition effects. The YOLOv4-Tiny deep learning model with added pretraining performs best on a customized test set. Compared to the best Map50 [mean average precision (AP)] model YOLOv5s, this article reduces the number of parameters by a factor of 13.4 and also maintains Map50 at 94.86%. The forward reasoning process that completes a single session on a CPU takes 3.52 s, but using Zynq accelerates this by 16 times, reducing the time to just 0.22 s and the power consumption to only 3.1 W. It can achieve high efficiency, low execution time, and low power consumption, and can be used in real time for FAIMS spectral recognition.