Research On Bionic Data Analysis For Electronic Noses

Electronic nose(e-nose)is an instrument consisting of a cross-sensitive gas sensor array for mixed odor detection and recognition.Compared with conventional gas analysis instruments,e-nose has been widely used in petroleum chemical industry,food inspection,medical diagnosis and other fields owing to its advantages of rapidity,portability and low price.Although e-nose imitates the principles of biological olfaction in structure and function to realize odor recognition,the existing e-nose data analysis methods generally contain many steps.There are a variety of approaches available in each step.To obtain the best detection performance,it is necessary to combine and optimize algorithms of different steps according to actual demands.This process often takes a lot of time and effort,which restricts the development of e-nose technology.As human understanding of biological mechanisms continues to deepen,bionic achievements are increasingly used in engineering fields,providing new ways and ideas to solve technical problems.This dissertation first introduces the traditional e-nose data processing framework,and then based on the bionic models,the e-nose data analysis is carried out from single platform to cross-platform.The main research contents are as follows:Based on the traditional e-nose data processing framework,a preliminary attempt has been made to identify odors.The traditional data analysis methods,which contain signal preprocessing,feature extraction,feature reduction and classification,are designed according to the gas sensors' response characteristics in different e-nose platforms.To validate the algorithm performance,the comparison experiments have been performed using seven kinds of Chinese liquors.The experiment results show that traditional data analysis methods can effectively solve the problem of multi-class classification and achieve high classification accuracy for different sampling approaches and platforms.To simplify the procedure of traditional e-nose data analysis methods,a signal preprocessing and feature analysis method using spiking olfactory bulb model is proposed.According to the basic structure and characteristics of biological olfactory system,a spiking olfactory bulb model including spiking olfactory receptor and primary olfactory bulb is established.Inspired by the information processing mechanism of biological neurons,the gas sensor responses are transduced into spike time sequences using receptive fields in spiking olfactory receptor neurons.The olfactory bulb model,consisting of mitral cells and granule cells,mimics the basic structure of biological olfactory bulb.The mitral cells receive the spike sequences generated by spiking olfactory receptor neurons and then transmit them to granule cells for further analysis.The bio-inspired method does not need filtering,denoising and feature reduction.It not only simplifies the processing procedure of traditional methods,but also obtains higher recognition accuracy than traditional ones.To improve the generality of e-nose data analysis process,a data analysis method based on spiking cortex model is proposed,and the spiking olfactory cortex model including improved spiking olfactory receptor,primary olfactory bulb and olfactory cortex is established.Firstly,the activation function of spiking olfactory receptor neurons is improved to enhance encoding efficiency.In addition,the primary olfactory bulb model is modified according to the structure of biological olfactory bulb.Then an advanced olfactory bulb model including the principal neural circuits of olfactory bulb is obtained to realize the extraction and processing of odor information.It is more in line with the physiological characteristics of biological olfactory bulb.Based on a learning rule with biological characteristics,an olfactory cortex model is established to perform learning and classification for olfactory bulb model's output.This method not only has the advantages of bionic olfactory model without filtering,denoising and feature reduction,but also can accomplish recognition tasks through automatic learning.The experimental results under different conditions show that the spiking cortex model has better classification performance and robustness in data analysis compared with the spiking olfactory bulb model.To solve the problem that the existing e-nose data analysis methods cannot handle cross-platform data,a cross-platform data analysis method based on time series visualization is proposed.Firstly,the visualization of time series is performed for gas sensor responses so that the sample data of different platforms can be transformed into a unified form of combination images.Subsequently,a small convolution neural network model is built according to the characteristics of e-nose data.The overfitting problem caused by small sample data is eliminated by using various strategies,and the network can automatically learn deep features for the classification of cross-platform data.The experimental results of three platforms show that the proposed method is feasible for different sensor arrays and sampling approaches and has high classification performance for all platforms.This realizes the general e-nose data analysis,which is expected to promote the popularization and application of e-nose technology.