Study On Spectral Methodologies For Quality Assessment Of Infant Formula Powder

Infant formula powder is the only source of nutrition for non-breastfeeding infants,and its quality and safety are closely related to infant health.Traditional detection methods can not meet the severe demand for dairy product quality and safety supervision due to the expensive instruments and complex operation.Therefore,the development of novel detection methods for screening the quality and safety of milk powder samples is extremely necessary,which also represents the research forefront in the field of biomedical engineering.In this thesis,based on laser-induced breakdown spectroscopy(LIBS)and Raman hyperspectral technology,a series of new spectral data processing methods,such as data-driven strategy,non-targeted screening strategy,deep convolution neural networkdriven spectral reconstruction,have been developed to detect the quality and safety of infant formula powder according to the spectral characteristics of infant formula powder.The main contents of this thesis are as follows:(1)A data-driven laser-induced breakdown spectroscopy quantitative detection method(DD-LIBS)was proposed,which can effectively meet the needs of rapid quantitative analysis of key elements in infant formula powder.Aiming at the LIBS spectral interference caused by the plasma complex formation mechanism and the influencing factors,the data driving strategy is proposed innovatively.This strategy makes full use of the oversampling decomposition capability of high-density wavelet transform(HDWT)and the variable extraction capability of the random frog algorithm,which is combined to extract the intrinsic features accurately in the approximate shiftinvariant high-density wavelet space.A quantitative calibration model is thus built.It effectively improves the accuracy of LIBS technology for the quantitative analysis of key elements in milk powder.(2)A portable Raman hyperspectral imaging system was established to effectively meet the needs of accurate quantitative measurement of heterogeneous solid systems such as milk powder.Raman hyperspectral imaging system is capable of collecting sample information in a large area without losing sample details,possessing a broad future in biomedical engineering.On this basis,the Raman spectra pretreatment method together with the quantitative and visual detection method of common adulterants in milk powder were proposed,which provides a promising tool for the safety detection of milk powder samples.(3)Two non-targeted detection methods based on the moving window spectral angle measure(MWSAM)and the protein concentration probe(PCP)were developed to accurately extract the geometric features and regression spatial features of the Raman hyperspectral data,respectively.The Raman hyperspectral data of typical and wellcontrolled genuine milk powder samples were used as reference databases.The MWSAM method was developed based on the Raman hyperspectral characteristics and geometric features of genuine milk powder samples.The authenticity of milk powder samples was identified by comparing the similarity of the distribution of spectra angle features between the unknown sample and the genuine samples.The PCP method was further developed based on the regression spatial features of the genuine milk powder Raman hyperspectral analysis model.By mapping spectral variables to protein model space,the unknown adulteration was uniformly reflected in the deviation and disturbance in the protein regression model.These two methods complement and correct each other,which is expected to be capable of identifying the unknown adulterants in milk powder samples.(4)A deep convolution neural network-driven Raman hyperspectral reconstruction method is developed to improve the speed of Raman hyperspectral imaging significantly.This method takes the long-time integral spectrum as the ground truth,whose information is used to train the spectral reconstruction deep convolution neural network(SRDCNN).The SRDCNN extracts the intrinsic spectral features from the short-time integral spectrum to reconstruct the high signal-to-noise ratio spectral data.The satisfactory results illustrate that the SRDCNN improves the Raman hyperspectral acquisition efficiency through shortening the integration time,providing a new means for denoising and reconstruction of other one-dimensional data.