Research On Application Of Deep Learning In Near Infrared Spectroscopy

Analytical technique for near-infrared spectroscopy(NIRS)is commonly used to perform qualitative and quantitative analysis of hydrogen-containing compounds on analytes.Due to the properties of fast(on-line),non-polluting and non-destructive,NIRS has been widely used in the field of food,agriculture,bio-chemicals and medicine.The development of convenient NIRS instruments has brought this technique from industrial application to the public life,which greatly enriches the NIRS database.Traditional analytical modeling methods cannot meet the requirements of analyzing large amounts of data.Therefore,the "data-sensitive" deep learning method has gradually become a research hotspot in NIRS.In order to solve the problem of poor performance of deep learning methods when modeling data is insufficient,this paper studies the application of deep learning in NIRS from the perspective of improving the performance of deep learning algorithms and using deep learning methods to expand the scale of modeling data.(1)Aiming at solving the problem that the single deep learning model does not predict well when the amount of modeling data is insufficient,this paper proposes to integrate multiple one-dimensional convolution neural networks(1-d CNN)and use Negative correlation learning(NCL)to train these 1-d CNN methods(CNN＿NCL)in parallel to construct a NIRS quantitative analysis model.The experimental results show that the prediction ability of CNN＿NCL is always better than single 1-d CNN.And the larger the modeling data,the more obvious the advantage,which further expands the deep learning method in the "large NIRS database" modeling environment.(2)In order to enrich the modeling sample set,this paper proposes to use the Boundary Equilibrium Generation Adversarial Network(BEGAN)to generate virtual spectra that match the real sample distribution from the existing NIRS sample sets.At the same time,the virtual spectrum is successfully applied to the fusion algorithm,which improves the performance of it.The research results in this paper enrich the application of deep learning in NIRS,and provide new ideas for deeper and wider applications in the future.