ABSTRACT
In the present paper, a new model-based method was proposed for temperature prediction and correction. First, a temperature prediction model was obtained from training samples; then, the temperature of test samples were predicted; and finally, the correction model was used to reduce the nonlinear effects of spectra from temperature variations. Two experiments were used to verify the proposed method, including a water-ethanol mixture experiment and a ternary mixture experiment. The results show that, compared with classic method such as continuous piecewise direct standardization (CPDS), our method is efficient for temperature correction. Furthermore, the temperatures of test samples are not necessary in the proposed method, making it easier to use in real applications.
ABSTRACT
An FTIR spectrum fitting algorithm based on continuous wavelet transform is proposed. In calculating the factor of difference spectrum, the algorithm takes into account both the original spectrum and its continuous wavelet transformed spectra, which effectively overcomes the problem of reference peak selection and manual factor selection in most commercial software. The detailed discussions on wavelet scale, order and basis are included. The spectral fitting is performed on six wavelet basis functions and the obtained scale factor is used to quantify the content of liquor, and the corresponding mean absolute error ranges from 0.047 to 0.072, and the standard deviation ranges from 0.056 to 0.091. Experimental results show that the CWT combined with least squares fitting provides an accurate and reliable new method for FTIR spectral subtraction.
ABSTRACT
Chinese liquor is a complex mixture and contains a large amount of microconstituents, which affects the quality and flavor of liquor. In order to discriminate liquor flavors rapidly, the spectra of liquors were obtained by FTIR and employed as the input patterns of pattern classification algorithms, then liquor flavor discrimination models were built. This paper introduces liquor flavor pattern recognition algorithms comprehensively and systematically for the first time, and the algorithms contain statistical classifications (linear discriminant function, quadratic discriminant function, regularized discriminant analysis, and K nearest neighbor), prototype learning algorithm (learning vector quantization), support vector machine and adaboost algorithm. Experimental results show that the liquor flavor classification algorithms demonstrate good performance and achieve high accuracy, recognition rate and rejection rate.
