Application Of Meta-Heuristic Algorithm In Spectral Non-linear Modeling Optimization

Fourier transform near-infrared(FT-NIR)spectroscopy is a fast analysis technology that can integrate information detection in the fields of medicine,agriculture,food,and environmental science.However,the FT-NIR spectral data of the complex system contains comprehensive response information of multiple components at the same time,and the information overlaps seriously.The study of effective statistics/chemometrics methods and the accurate extraction of effective information of the components to be measured from spectral data are of great significance to the development of intelligent rapid detection and informationization in the fields of medicine and other fields.Multivariate calibration methods can be used to extract relevant information from different types of spectral data to obtain the analytical concentration or properties of complex samples.Using linear methods to model the data to be measured requires a linear relationship between the spectral data of the sample to be measured and the chemical reference value.The content of a single chemical component of a complex research object is not linearly related to the spectral data of the sample to be tested.Linear methods are difficult to obtain better modeling capabilities in the modeling and analysis of complex research objects.Therefore,classical linear methods are unsuitable for modeling,thus more complex methods should be sought in order to non-linearity modeling.The non-linear methods have strong nonlinear transformation ability,which can solve the nonlinear relationship between the spectral data and the chemical reference values,and avoid processing the nonlinear relationship into an approximate linear relationship.Although non-linear methods can solve the problem of non-linear relationship of the actual spectral data,there are some modeling parameters in the non-linear methods,and the different parameter values have a great impact on the prediction results.The meta-heuristic algorithms are selected to deeply optimize the modeling parameters of the non-linear methods to improve the predictive ability of the model.Therefore,the non-linear methods selected in this paper include least squares support vector machine(LSSVM),kernel partial least squares(KPLS)and kernel principal component analysis(KPCA).The meta-heuristic algorithms select genetic algorithm(GA),differential evolution(DE),grey wolf optimizer(GWO)and firefly algorithm(FA).Taking the rapid detection of the berberine components of Coptidis Rhizoma by FT-NIR spectroscopy as an example,the berberine content in Coptidis Rhizoma was quantitatively analyzed through the combined model of meta-heuristic algorithm and non-linear method,and the evaluation indicators of each model were compared.As a result,the prediction effect of the DE-LSSVM model will be better than the other combined models.The model evaluation indicators RMSET,RSD_T,and R_T values of the DE-LSSVM training set are 0.044(%),10.228,and 0.997,respectively.At the same time,the evaluation indicators RMSEP,RSD_P,and R_P values of prediction set are 0.157(%),7.656,and 0.944,respectively.The results show that modeling parameters of non-linear methods are optimized by using meta-heuristic algorithms,and then debugged parameter sets are applied to the non-linear methods.The combined model of meta-heuristic algorithms and non-linear methods can better solve the non-linear relationship between spectral data and chemical reference values,and improve the prediction accuracy and robustness of the prediction models.In FT-NIR spectral analysis,the combined model has potential prospects in the rapid detection of other test samples.