Characteristics Variable Selection of NIR Based on L1/2 Regulation

L
_P
norm (P<; l) has the feature of sparsity. In view of the sparsity of the characteristic band in the near infrared spectrum, a feature extraction algorithm based on L
_1/2
regular is proposed in this paper. According to the nonconvexity of L
_1/2
norm, an iterative algorithm is presented. The L
_1/2
regular is transformed into a series of L
₁
regular for iteration calculation. In this paper, based on the near infrared spectral data, L
_1/2
regular, PCR and PLS algorithm are used to detect oil compositions. And then the PLS algorithm and PCR algorithm are compared, the result of which shows that PCR and PLS algorithm have the similar effect and there will be a phenomenon of over fitting, as the number of characteristic band increase. Besides, the modeling sets error decreases and the prediction sets error increases. With L
_P
algorithm, both modeling sets and predictive sets error can be small, and extract the characteristic band effectively. The experiments show that the L
_1/2
regular has a good application value for the extraction of characteristic band of spectrum.