Research On Deep Learning Based Chinese Liquors Recognition Using Electronic Nose Systems

Electronic nose(e-nose)is a detection instrument inspired by biological olfactory systems,which consists of a sensor array,a signal preprocessing module and a pattern recognition system.Since e-nose was invented,it has been widely used in food safety,environmental monitoring and disease diagnosis.The data processing is an important part of e-nose technology,and the traditional data processing procedure of an e-nose mainly consists of several steps,i.e.,pre-processing,feature extraction(feature generation and reduction)and classification.Each step has plenty of optional methods or algorithms.Because there is no uniform rule,it is complicated and empirical for choosing appropriate methods.Moreover,traditional methods may not be universal for different kinds of e-nose data.Aiming at solving the problems of the traditional data processing in e-noses,this thesis carries out the research on the data processing method based on deep learning,and combines this method with the self-designed e-nose system,to classify different brands of Chinese liquors.The main research is presented as follows:A stacked sparse auto-encoders(SSAE)is proposed for feature learning of e-nose data.This method is a deep learning model,which adopts a greedy layer-wise unsupervised algorithm to pre-train,and learns the features from the original e-nose data automatically instead of the manual features of traditional methods.The back propagation neural network(BPNN),is constructed according to the structure of SSAE network and initialized with the parameters obtained in the pre-training of SSAE.Then the whole network is fine-tuned using the back propagation algorithm with the labeled data.The Softmax regression function is added to the output layer of BPNN to realize the classification of different kinds of Chinese liquors.In order to obtain the optimal classification effect of the SSAE-BPNN,a large number of experiments are designed to optimize the network structure and the parameters.The performances of the classifiers are evaluated by the cross validation approach,and the classification accuracy is chosen as the evaluation index.Firstly,the network structure of SSAE-BPNN is designed through tests,and then a simple network is built with two hidden layers to alleviate the over-fitting problems effectively.Secondly,the appropriate hyper parameters and training iteration times of SSAE-BPNN are selected through experiments.To verify the advantages of SSAE-BPNN algorithm,different contrast experiments are designed.Firstly,stacked auto-encoders(SAE)and SSAE are compared,and they are connected with the same BPNN classifier after feature learning.Secondly,support vector machines(SVM)and BPNN classifiers are compared,and they use the same SSAE to learn the features.The results show that the classification effect of SSAE-BPNN is superior to that of SAE-BPNN and SSAE-SVM.Finally,the deep learning methods are compared with the traditional methods,and the deep learning methods not only obtain higher accuracy,but also simplify the procedure of data processing.