Particle swarm optimization-based wavelet packet regression for multivariate analysis of near-infrared spectroscopy

Near infrared (NIR) spectroscopy, combined with multivariate calibration method, is a very important issue for qualitative and quantitative application. In pursuit of this aim, a new hybrid algorithm (PSO-WPLS) was proposed for multivariate regression model development. At first, wavelet packet transform (WPT) algorithm and its reconstruction algorithm are used to split the collected spectra into different frequency components. Then, to take advantages of multiscale property of NIR spectra, the useful WPT components are selected by the particle swarm optimization (PSO) algorithm coupled with a fitness function of prediction error. At last, each selected WPT components are introduced to regression models to develop a series of sub-models. The PSO-WPT model can be constructed through the involvement of all sub-models characterized by a series of weighted regression coefficient. To validate this algorithm, it was used to measure the oil concentration of corn samples. Compared with the conventional WPLS algorithm, the PSO-WPLS algorithm can significantly improve the quality of regression model with the prediction errors decreasing by up to 72.5%, meaning that it is a potential way for developing multivariate model with high precision.