High-Performance Sensing: Application of FPGA-based Winograd CNN Accelerator in Electronic Nose Systems

Electronic nose technology has been widely used in various scenarios such as industrial, military, food and environmental, and its recognition accelerator is responsible for completing the detection and identification of gases. Therefore, the application of advanced artificial intelligence algorithms to recognition accelerators is a common practice among researchers nowadays. However, neural networks with high recognition rates tend to have complex structures and a large number of parameters, which exacerbates the difficulty of deployment. To address this challenge, this paper first analyses and processes the detected gas data and proposes a hardware-friendly lightweight gas classification network, and then, based on this network, designs an FPGA-based Winograd CNN accelerator through the study of hardware implementation of Convolutional Neural Networks (CNN). The Winograd algorithm is used to reduce the number of multiplications used in the convolutional operation so as to reduce the consumption of hardware computing resources by the accelerator. Finally, the designed gas classification network is deployed on FPGA to complete the implementation of the recognition acceleration engine in the electronic nose system. The detection and recognition experiments on the laboratory gas dataset demonstrated good performance in terms of 99.5% recognition rate and only 2.94W power consumption of the accelerator, respectively.