Research Of Partial Least Squares Regression Algorithm Based On Optimal Selection Of Latent Variables

Near infrared spectroscopy signal represents the composition or concentration of the substance,the NIR analysis technique because of its rapid,nondestructive,low cost,etc.,in the food,agricultural products,pharmaceutical and biotech industry are widely used.Least Squares Regression Partial(PLSR)is a kind of commonly used near-infrared spectral correction model,which is used to predict the concentration of samples or the content of components.The PLSR based cross validation is used to select the latent variable,while the latent variable is sometimes important for the determination of certain components.In this paper,the latent variable optimization method is presented,and based on latent variable to optimize the selection of partial least squares regression(LOPLSR),mainly used for the concentration or composition of near infrared spectral data to predict.The main innovations of this paper are as follows:Firstly,the optimization of the latent variable is presented,and the PLSR algorithm based on the method is given.The basic idea of the algorithm is to find the positive correlation latent variables,and remove the negative correlation latent variables,so as to improve the prediction accuracy.The training sample set is divided into a training set and a calibration set,using the original PLSR algorithm obtained training set the latent variables,projection matrix,load,the prediction accuracy decreased the number of latent variables,through the projection matrix obtained continuously updated,calibrating set score.According to the load of the training set,the score of the calibration set is predicted by the material concentration or the composition.Respectively by two different infrared spectral data of grass and orangejuice PLSR and LOPLSR comparative experiments,the optimal latent variables selected by cross validation method to get.Experiments show that,improved LOPLSR algorithm than the original PLSR algorithm to predict the result more accurate and minimum increased by 24.9%.Secondly,the method of setting the weight of the latent variable in the algorithm is given.Algorithm in this paper according to the obtained weight matrix constantly updated to obtain calibration set of latent variables,and finally to make the prediction accuracy decreased number of latent variables of training load and calibration set scores were multiplied by the weight 0,otherwise 1 multiplied by their weights,until the operation is completed,you can get a right value of the diagonal matrix.In the process,the potential variable of the variation of the forecast results is removed,so the accuracy of the forecast can be improved to some extent.Finally,a method for the detection of the concentration of the peony and the method for the detection of the meat components is presented.By analyzing examples of meat ingredients and paeoniflorin concentration detection results show that LOPLSR algorithm is compared with the traditional PLSR algorithm has better prediction accuracy,indicating that the algorithm LOPLSR in practical application to detection of meat ingredients and paeoniflorin concentration,but also has a better detection performance than PLSR algorithm.