Determination of Caffeine Content in Coffee Beans Based on Laser Induced Breakdown Spectroscopy
Song Kun-lin,Zhang Chu,Peng Ji-yu,Ye Lan-han,Liu Fei,He Yong
DOI: https://doi.org/10.3964/j.issn.1000-0593(2017)07-2199-06
2017-01-01
Spectroscopy and spectral analysis
Abstract:The feasibility of fast detection of caffeine content in coffee beans based on laser induced breakdown spectroscopy (LIES) combined with chemometrics methods was studied. The coffee beans were grinded. 0.5 g of powerd material was transformed into a disk by manual tableting machine. 60 disks of coffee bean material were prepared for LIBS data acquiring. The samples were pretreated by acid wet digestion, and actual caffeine content of each sample was obtained by automic absorption spectrometer (AAS). Baseline correction was applied on the original spectral data to eliminate the negative values. Wavelet transform (WT) was used to reduce the noise, wavelet basis function is Daubechies 5 (db5) and decomposition level is 10. Normalization were employed to deal with variations caused by matrix effects and experimental conditions. The partial least squares (PLS) model on full data appeared to over-fitting. Regression coefficients and principal component analysis (PCA) were used to select characteristic variables, respectively. PLS models and back propagation (BP) neural network model were built by the variables selected. In the PLS model on the variables selected by regression coefficients, correlation coefficient of calibration set (R-c) was 0.96, correlation coefficient of prediction set (R-p) was 0.91. In the PLS model on the variables selected by PCA, R-c=0.94, R-p=0.90. In the BP neural network model on variables selected by PCA, R-c=0.96, R-p=0.96. The characteristic variables selected by two methods correspond to C, H, O, N, Na, Mg, Ca, Fe, Mn. The PLS models on the variables selected by regression coefficients and PCA both performed well on prediction samples. It demonstrated the certain relationship existing between the elements and caffeine content and the selected variables were effective. But the precise relationship bwtween C, H, O, N, Na, Mg, Ca, Fe, Mn and caffeine content needs further study. The BP neural network model on variables selected by PCA performed better than the PLS model, which demonstrated the selected variables were suitable for different modeling methods. The study showed LIBS could be applied to fast determination of caffeine content within coffee bean combined with chemometrics methods. The method of caffeine content detection presented by this study is innovative.