Research On Electronic Nose Odor Identification Based On ZYNQ

Electronic nose(e-nose)is an instrument composed of the sensor array and appropriate odor identification algorithm.It has been widely used in various industries related to odor detection,such as food safety,environmental monitoring,space shuttle,and medical diagnosis,and it plays an important role.Designing a more efficient and low-cost integrated e-nose system is one of the research hotspots of odor identification.In the design of the integrated enose,two independent hardware and software modules,the sensor array module and the recognition algorithm module of the ordinary e-nose are integrated into one.Compared with the ordinary e-nose system,this design has obvious advantages.It avoids using additional computer hardware for the recognition algorithm,reduces the complexity of the electronic nose system,improves mobility,and reduces data transmission,enabling the e-nose to perform real-time odor identification.However,the current integrated e-nose system is limited by the hardware system.There are problems such as the low odor identification accuracy of the integrated e-nose system with a simple algorithm structure and the slow running speed of the integrated e-nose system with a complex algorithm structure.This paper proposes implementing a deep neural network for odor identification in a low-cost Field-Programmable Gate Array(FPGA).First,a lightweight deep learning odor identification with a depthwise separable convolutional neural network V2(OI-DSCNNV2)is proposed to reduce parameters and accelerate hardware implementation performance.Secondly,OI-DSCNNV2 is implemented in the Zynq-7020 So C chip based on the saturation and flooring-pearson correlation coefficient(SF-PCC)quantization scheme.Finally,the OI-DSCNNV2 model was run on the Chinese herbal medicine dataset,and the simulation experiment and hardware implementation on FPGA proved its effectiveness.The main research content of this article is summarized as the following three points:(1)Research on odor identification method based on depthwise separable convolution.This paper proposes a lightweight deep learning odor recognition model called OI-DSCNN and improved it to OI-DSCNNV2 odor model.The OI-DSCNNV2 dramatically reduces the number of the deep learning algorithm parameters in the odor identification process while balancing the speed and accuracy of the odor identification algorithm.Through experiments,OI-DSCNNV2 is applied to the dataset of Chinese medicinal materials.Experimental results show that,compared with complex deep learning models such as CNN,CNN-SVM and OLCE,OI-DSCNNV2 has a simpler structure and maintains accuracy.Compared to odor identification models with simple algorithm structures such as LDA,MLP,DT and PCA+LDA,and other machine learning algorithms,OI-DSCNNV2 has higher accuracy.(2)Research on FPGA-oriented low-precision quantization methods.In order that the OIDSCNNV2 model can adapt to the hardware structure of FPGA and run at high speed in Arty Z7-7020,SF-KL and SF-PCC quantification schemes are proposed,and the two quantification schemes are compared.The SF-KL and SF-PCC quantization schemes use saturation cutoff and round-down methods to quantify the output of the hidden layer of the model,and use KL divergence and Pearson correlation coefficient to select the optimal cutoff position scheme for each layer.Experimental results show that SF-PCC is better than SF-KL,and this method reduces FPGA resource consumption and maintains the accuracy of OI-DSCNNV2 model for odor recognition.(3)Based on ZYNQ accelerated OI-DSCNNV2 model construction.This paper designs and optimizes the architecture and modules of the OI-DSCNNV2 model in the low-cost development board Arty Z7-7020 based on the ZYNQ architecture.By studying the model structure of OI-DSCNNV2,the hardware in Arty Z7-7020 is reasonably allocated.Resources.The development of an electronic nose odor recognition module based on Arty Z7-7020 speeds up the speed of odor recognition and solves the problem of slow operation of the integrated electronic nose system due to the complex algorithm structure.