Wavelength Selection Based On Affinity Propagation Algorithm

Spectral analysis can predict the components of a sample by using its spectrum data and a prediction model.The performance of the prediction largely depends on the input wavelengths into the prediction model.To decrease the complexity of the model and increase the prediction performance,it is necessary to conduct wavelength selection before developing the prediction model so as to reduce the dimensionality of the spectrum data,remove noise and redundant information form the spectrum data.Among various wavelength selection methods,the combination of Physarum Polycephalum Network(PN)and Genetic Algorithm-Partial Least Squares(GA-PLS)is a new and effective wavelength selection method.However,the PN requires that the entire spectral range must be divided into the equal multiple sub-spectral bands.This requirement will limit the variable selection capability and the prediction performance of the model.In order to solve this problem,this thesis proposed a wavelength selection algorithm based on Affinity Propagation(AP)combined with Physarum Polycephalum Network(PN)and Genetic Algorithm-Partial Least Squares(GA-PLS).The idea of the algorithm is that transforming the spectrum division problem into a clustering problem,and using the Affinity Propagation(AP)algorithm to find the optimized number of sub-spectral bands and the number of wavelengths in each sub-spectral band according to the feature of the spectrum.And then using PN to select the candidate variables.Finally,the candidate variables were selected by GA-PLS.In this thesis,the AP-PN-GA-PLS was tested on two spectral databases,and was compared with full spectrum feature variables,Interval Partial Least Squares(IPLS),Successive Projections Algorithm(SPA),GA-PLS,AP-GA-PLS,PN-GA-PLS.The experimental results showed that AP-PN-GA-PLS could achieve the best prediction performance among the seven wavelength selection algorithms with the minimum number of wavelength selection variables.Thus,AP-PN-GA-PLS can minimize the complexity of the model without compromising prediction performance.