Feature selecting based on fourier series fitting

In spectral quantitative analysis, the accuracy and complexity of the designed prediction model will be negatively affected by enormous data volume and noise of the original spectrum. This paper presents a dimensionality reduction and noise decreasing method for original spectrum analysis based on Fourier series fitting (FSF). By extracting features using FSF, the original spectrum data will be mapped into Fourier series, and a regression prediction model using partial least squares (PLS) is established. The experimental analysis suggests that FSF method, compared with PLS, discrete fourier transform (DFT), artificial neural networks (ANNs) and genetic algorithm with PLS (GA-PLS), can produce better results considering the running time, the number of input variables and prediction accuracy.