Co-Design of Lightweight Neural Network and Low-Power Edge Computing Architecture for Intelligent Electronic Nose System

The powerful feature extraction capabilities of deep learning can improve the recognition performance of the electronic nose system effectively. However, there is a conflict between the computing load of traditional deep learning algorithms and the limited computing resources of edge devices, which severely limits the application of the electronic nose on edge devices. In the face of these problems, previous studies have focused on algorithm optimization by introducing model compression methods. However, two challenges remain. Specifically, these are redundant parameters and high power consumption. Therefore, in order to solve the above problems, this article first establishes a new industrial gas dataset, which contains ten common harmful industrial gases and 6000 pieces of corresponding 2-D gas data. Then, we propose a lightweight and effective model called focused residual convolution neural network (FRCNN). The goal of FRCNN is to reduce the number of parameters and balance computational efficiency with accuracy. Compared with the traditional gas recognition algorithms, FRCNN has been significantly improved. In addition, in order to realize an edge-type electronic nose with low power consumption and low resources, we construct a reconfigurable convolution acceleration engine in pipeline form based on Winograd algorithm and design a hybrid quantization training method to reduce the field-programmable gate array (FPGA) deployment resources greatly. Finally, sufficient experiments and comparisons are conducted to verify the effectiveness of the proposed approach. The results show that when the recognition accuracy of FRCNN reaches 99.71%, the number of network parameters decreases to 1/4. After adopting the quantization algorithm, the loss of recognition accuracy is less than 2%, and the consumption of hardware resources is reduced to one-fourth of the original. Besides, when the acceleration engine is deployed on the xc7z020clg400 chip, the resource utilization is less than 10%, and the recognition speed and energy efficiency can reach 100 frames per second (fps) and $10\times $ , respectively.