Research On Neural Network Adaptive Optimization Method For Spectral Quantitative Analysis

Quantitative analysis is an important statistical modeling method.It establishes a mathematical model based on statistic,and uses this model to calculate various values of the analysis object.Quantitative analysis has been applied in many research fields to take into account the effectiveness of this modeling method.Among them,the near-infrared(NIR)spectroscopy analysis technology is a combination of statistics,analytical chemistry,and other cross-disciplines,and it is an efficient and non-destructive analysis technology.The ability of this technology to obtain a satisfactory analysis effect largely depends on the performance of the statistical model used.Therefore,how to better design a reasonable quantitative analysis model in the NIR analysis technology is a problem worthy of study.Artificial neural network is a comprehensive and nonlinear mathematical model suitable for spectral analysis.Therefore,this type of model is used to apply to NIR analysis,and proposes an adaptive cuckoo search neural network(A-CSNN)model on the basis of artificial neural network.It is a neural network model that uses cuckoo search to optimize parameters such as weights and deviations.At the same time,an adaptive function is added to adjust the parameters in the calculation process to further improve the accuracy of the model.Whether the A-CSNN model can be applied to NIR spectroscopy analysis still needs specific experimental analysis.The experimental data is selected from fishmeal as NIR spectroscopy data,and the protein content of sample is determined by chemical determination method as reference value.After that,sample outliers are screened and sample features are reduced.Cross-validation is used to test the performance of the model,and calculate the mean value and standard deviation of the results.Subsequently,the experiment divided the fishmeal samples into calibration set and test set according to specific sample division methods,and established models to predict the results.In order to facilitate the observation of model effect,the back propagation neural network(BPNN)model,the adaptive back propagation neural network(A-BPNN)model,and the cuckoo search neural network(CSNN)model are selected as the comparison models of ACSNN.In the analysis of cross-validation,the mean value of root mean square error of A-CSNN model is 2.054%,and the mean value of correlation coefficient is 0.877;the standard deviation of root mean square error is 0.177%,and the standard deviation of correlation coefficient is 0.018.In different sample partitioning methods combined with neural network model experiments,the adaptive hybrid cuckoo tabu search(A-HCTS)partitioning method combined with the A-CSNN model has the best effect.The root mean square error of the calibration set is 1.280%,and the correlation coefficient value is 0.958;the root mean square error of the test set is 1.359%,and the correlation coefficient value is 0.934.From the experimental results,it can be seen that the A-CSNN has significant advantages in NIR spectroscopy,and its results are better than other three models.Therefore,it can be explained that the A-CSNN is a feasible and effective statistic algorithm in the NIR spectroscopy analysis.