Research On Abnormal Data Processing And Feature Extraction Method Of The Electronic Nose

Nowadays,environmental problems and food safety problems have become the focus of the society.Therefore,the objective and accurate identification of gas is closely related to people's lives.The electronic nose is a device imitating the animal olfactory mechanism to realize qualitative and quantitative identification of gas.In recent years,the research results of the electronic nose are emerging and relevant technology is developing rapidly.The gas sensor of the electronic nose is unstable and vulnerable to environmental factors.So,the emergence of abnormal data is inevitable,which affects the effect of subsequent feature extraction method.At present,the feature extraction method of the electronic nose can be divided into local feature extraction method and integral feature extraction method.The local feature extraction method considers the local information of the curve,which extracts the value of the response curve at a moment.The integral feature extraction method considers the whole information of the curve,and deals directly with the whole response curve,mostly adopting the curve fitting method.In most studies,only one feature extraction method is considered.The local feature extraction method mainly relies on the geometric features of the response curve.However,apart from the gas,the response of the electronic nose's gas sensor is affected by many factors.The curve fitting method depends on the fitting model and is an approximate method.Therefore,both methods all lack accuracy and reduce the recognition rate of the electronic nose.According to the above problems,this thesis presents a abnormal data processing algorithm using k-means clustering algorithm combined with weight to optimize t test criterion and a feature extraction method combining local and integral features.The former uses k-means clustering algorithm combined with weight to analyze the sample data and extract the modified data.It replaces abnormal data with modified data after t test criterion detects abnormal data.Because the response(curve)of the electronic nose's gas sensor is divided into several different stages,the data is processed in group to enhance the robustness of the algorithm.The latter combines local feature extraction and integral feature extraction to make sure that the features of the response curve of the electronic nose's gas sensor are extracted comprehensively and accurately.This thesis proposes that the feature extraction method combining local and integral features uses Fourier transform to transform the original data in the time domain into the frequency domain aiming to extract the peak value of the curve in the frequency domain as the integral feature,which avoids the inaccuracy of model in the curve fitting method.216,000 data of C2H5 OH and NH3 samples were obtained in the experiment.Simulation and verification have been done on Matlab.The result shows that the algorithm using k-means clustering algorithm combined with weight to optimize t test criterion can detect and correct abnormal data accurately and effectively.What's more,the error between the corrected data and the normal data is within 0.5 and the relative error is within 4.63%.Comparative study on the local feature extraction method,the curve fitting method and the feature extraction method combining local and integral features has been done by simulating.Finally,iteration steps of the support vector machines(SVM)are 32,53 and 28 respectively,and the recognition rates are 85%,70% and 95% respectively.It is proved that the feature extraction method combining local and integral features presented in this thesis can improve the recognition rate of the electronic nose.