Near-infrared Spectral Analysis Based On Deep Learning

The rapid development of near-infrared spectroscopy benefits from the application of chemometrics.With the increasing complexity of spectral data,the traditional model is difficult to meet the demand due to weak adaptability,so that the near infrared spectroscopy technology is facing a huge challenge.Therefore,the establishment of robust and reliable model by chemometrics methods is the key to solve the problem.In this paper,a near infrared spectral analysis algorithm based on random hidden deep belief network(DBN)is proposed,which is applied to the analysis of the content of cotton polyester blended fabric and the identification of ripe pears.The main research contents and conclusions are as follows:(1)In the sample selection,the Mahalanobis distance method is often used to eliminate the abnormal samples to prevent all aspects of subjective and objective factors to trigger the existence of abnormal spectral data in modeling data.Due to the complexity of original spectral data and the doping of irrelevant information,the influence of pretreatment methods such as SG smoothing,MSC and SNV on Modeling and analysis is compared in the pre-processing of spectra.Finally,the wavelet transform is used to extract the original spectrum.The root mean square error of wavelet reconstruction is used to select the wavelet function and decomposition layers,aimed to extract characteristic data,spectral noise removal,highlighting the true spectral information.(2)In the modeling and analysis,the near infrared analysis method based on random concealment DBN is proposed to solve the problem of weak adaptability of the model.The random hiddent algorithm was applied to the adjustment stage of parameters of deep belief network.Some network weights are not updated,which can effectively improve the overfitting phenomenon,improve the modeling speed,and get more stable model,which is conducive to the transfer and sharing of the model in practical application.(3)The random hidden DBN method was applied to quantitative and pattern recognition.The near infrared analysis method based on random concealment DBN was used to analysis the component content of cotton polyester blended.Thecorrelation coefficient of prediction of cotton and polyester content were 0.9941 and0.9965,the predicted standard deviation were 0.0253 and 0.0535.Compared with the traditional modeling methods(artificial neural network and partial least squares)and standard DBN,the results show that the effect of random hidden DBN model was significantly improved.The discriminant model baesd on random hidden DBN for judging the maturity of sand pear is established and the influence of the different pre-processing methods were discussed.As the result,WT was the best pre-processing method to establish a model which prediction accuracy was 95.8%.Similarly,random hidden DBN model were compared with the other three(PLSDA,DBN,autoencoder network)modeling method.The results show that the random hidden DBN model is better than the other three models,so it can be applied in pattern recognition.(4)NIR analysis software based on random hidden DBN was developed.The software contained two parts.One of them used for quantitative analysis and another used for pattern recognition.The Cotton polyester blend and sand pear samples were used to test the function of the software,and the predicted output of unknown samples was achieved.The software could be used in practical application.