Feature Reduction And Classifier Design Based On Electronic Nose System For Chinese Liquors Recognition

Chinese liquor culture is an important part of broad and profound Chinese culture.Wine and wine culture have almost penetrated every aspect of our social life,and become an indispensable part of people's lives.At the same time,liquor industry plays an important role in promoting national economic growth.However,with the development of liquor industry,it is noticeable that there are many problems in liquor-making industry,in which the fake liquors and adulterations are the most serious.The problems seriously affect the order of the liquor industry and bring harm to the consumers' health.At present,gas chromatography and mass spectrometry are major methods for analysis of Chinese liquors.But they are time/labor consuming and cannot realize rapid detection of Chinese liquors.Electronic nose(e-nose),as a new detection technology,is inspired by olfactory systems of mammals.It has been widely accepted as a quick,nondestructive,and effective method in many fields,such as food safety,medical filed,and environmental monitoring.Owing to the technical constraints,the e-nose system also has some disadvantages,such as the drift of the gas sensor array.Besides,the detection results of the e-nose system are easily affected by the environment.So,it is very important to study the data analysis method in order to improve the reliability of the e-nose system.This thesis focused on the feature dimension reduction algorithm and classifier design in the e-nose system.A more effective data processing method of electronic nose system for classification of Chinese liquors was put forward.The major work is presented as follows.A new hybrid feature reduction method,KECA-LDA,for e-noses is proposed.This method combining kernel entropy component analysis(KECA)with linear discriminant analysis(LDA)was used to reduce the dimensionality of the electronic nose data.The original data was firstly transformed into KECA space and then fisher criterion was applied to maximize the between class distance.Compared with KECA,the combination of KECA and LDA made the data more separable and easier to classify.Particle swarm optimization(PSO)algorithm was employed to optimize the kernel parameter in KECA.KECA algorithm used in this thesis is based on the Gauss kernel function.The performance of KECA is affected significantly by the kernel parameters.So selecting the most suitable kernel parameters according to a specific project is a crucial problem.PSO is easy to implement and has strong global convergence ability.The cross-validation accuracy was taken as the fitness function of PSO.In this thesis,extreme learning machine(ELM)was employed to recognize Chinese liquors with different brands.ELM is a fast learning algorithm for single-hidden layer feedforward neural networks and tends to reach the smallest training error.Unlike other traditional learning algorithms,ELM can randomly determine the input weights and biases of the hidden layer.The weights between the hidden layer and the output layer are the only parameters that need to be learned.In order to confirm the algorithm's effectiveness,comparison experiments are done using ELM and back propagation neural network(BPNN).