Study On Signal Denoising And Feature Extration Of Electronic Nose To Moldy Maize

In order to improve correct rate of discrimination results of moldy corn samples using an electronic nose,the influence of different features combination representation types of E-nose signals on the discrimination effect of complex samples was studied in depth.And a method based on Wilks ? statistics used to quantitatively evaluate the discriminant ability of feature vectors corresponding to multi-features combination is proposed.At the same time,considering the different selective characteristics of different gas sensors,an optimization method based on the elimination transform with pivoting of Wilks ? statistic used to select features representation of gas sensors is also proposed under the condition of multi-features combination representation mode.Moreover,in order to improve the prediction ability of mycotoxin content in moldy corn by electronic nose,under the condition of 5 features combination representation mode,the prediction model based on kernel Fisher Discriminant Analysis coupled with BP neural network is also proposed to improve the prediction ability of the electronic nose for moldy corn mycotoxin content.The specific research work is as follows:Firstly,integral value(INV),average differential value(ADV),relative steadystate average value(RSAV),variance value(VAR)and wavelet energy value(WEV)of every sensor signals for the E-nose were extracted as five kinds of features.Then,the discriminant ability of feature vectors corresponding to different feature analysis matrixes under the condition of single feature or multi-features combination representation mode was calculated by Wilks ? statistic,the results showed that the identification effect based on multi-features combination representation was better than that based on single feature representation and information of complex samples could be fully reflected with more features to characterize gas sensors signals,and with the increasing of the number of characteristic features,the discriminant ability of the corresponding feature vectors is further improved.Moreover,it also indicated that the combination method of multi-features was not a regular pattern,but the better features combination could be obtained by the Wilks ? statistic evaluation method.At the same time,under the condition of multi-features combination representation mode,by the way of feature selection proposed,optimization and selection of multi-features combination representation of gas sensors signals is carried out,The results show that the characteristic of different sensors is different under the condition of multi-featurescombination representation mode,which indicates the necessity of feature selection.In addition,in order to reveal the effectiveness of the evaluation method of the discriminant ability of features,Fisher Discriminant Analysis(FDA)is used to examine visually the identification results under the condition of single feature or multi-features combination representation mode.The results showed that no matter which feature representation mode,FDA results were consistent with the analysis results based on ? value,which shows that the evaluation method of discriminant ability of feature vectors is effective,and with the increasing of the number of characteristic features,the discriminant ability of the corresponding feature vectors is gradually improved.The correct rate of FDA for moldy corns under the condition of five features combination representation mode is 98%.This shows that the evaluation method is effective.Finally,with the help of BP neural network and kernel Fisher Discriminant Analysis(KFDA)coupled with BP neural network,the prediction models of Aflatoxin B1 content,Deoxynivalenol content(DON)and Zearalenone content were studied and built respectively.Correlation analysis and relative error analysis were carried out for the predicted and actual values,and the results showed that: when the BP neural network was used to build these prediction models and the correct sample number percentage of the forecast error that was within 5% was up to 85%;when BP neural network coupled with KFDA was used to build these prediction models and the correct sample number percentage of the forecast error that was within 0.6% was up to 100%.The fitting coefficient of the predicted and measured values of the two models was also increased from 0.95 up to 1.00.The research finding clearly shows that the method based on BP neural network coupled with KFDA can accurately predicts these moldy maize mycotoxin contents,and the detection ability of the E-nose for the moldy maize mycotoxin contents is significantly improved.The results of this paper can gain 4 conclusions: 1)Only with multi-features combination representation,can the complexity of the sample be reflected.2)The discriminant ability of feature vectors corresponding to multi-features combination could be evaluated by Wilks statistics.3)An optimization method based on the elimination transform with pivoting of Wilks ? statistic used to select features representation of gas sensors is proposed under the condition of multi-features combination representation mode.4)The method based on KFDA coupled with BP neural network is effective to predict moldy maize mycotoxin content.