Research On Pattern Recognition Of Near-Infrared Spectral Analysis Based On CNN And Ensemble Learning

With the development of near infrared spectroscopy(NIRS)and machine learning algorithms,more and more researchers have applied NIRS and machine learning algorithms to the pattern recognition of substances.NIRS is convenient,fast,efficient and nondestructive,and is used for the detection and identification of substances frequently.Machine learning is an important research direction of artificial intelligence.With the rapid development in recent years,especially the resurgence of neural network research,the research of machine learning has reached a new climax in the world.Convolution neural network(CNN)has strong pattern recognition ability,not only in two-dimensional images,but also in one-dimensional data.Ensemble learning can combine multiple weak classifiers into a strong classifier,and finally get the prediction results by voting the sub-models to be tested.In view of the limited model performance of CNN when the sample size is insufficient.In this thesis,we propose to improve the performance of the original CNN pattern recognition in the case of small samples by introducing transfer learning combined with CNN.The source domain model trained through a larger sample dataset has better model performance.Freezing and migrating the network parameters of the source domain model to the target domain model,initializing the full connection layer of the target domain model,and finally train the improved CNN model through migration learning.Through experiments and comparing the accuracy,accuracy,recall,and F1 values of CNN before and after the improvement.The experimental results show that the improved CNN scores higher than the original CNN in the four evaluation indicators.This indicates that the improved CNN model through migration learning has better model performance than the original CNN model in classifying pure naked oats and potato starch doped with 10%-50%concentration.In view of the fact that all sub-models of the traditional ensemble learning algorithm have the same voting weight and cannot play the advantages of the stronger sub-models.In this thesis,we propose Exponential-VS1-stack(VS1-stack)algorithm to adjust the voting weight of the ensemble learning sub-models,so that the voting weights of the ensemble learning sub-models are different.According to the sample classification performance of the sub-models,the voting weight of the model is calculated by VS1-stack algorithm.Through experiments and comparative analysis of the accuracy,accuracy,recall,and F1 value of integrated learning before and after improvement.Experimental results show that ensemble learning using VS1-stack algorithm to improve voting weight is superior to traditional ensemble learning with the same weight in four evaluation indicators.This indicates that VS1-stack ensemble learning algorithm has better model performance than traditional ensemble learning algorithms in classifying pure naked oats and potato starch doped with 10%-50%concentration.