| NIR reflects the information of molecular rotation and vibration,and the composition information of substances can be obtained from it.This method has the characteristics of zero pollution,no sample damage,simple operation,fast analysis,etc.Therefore,it has been widely used in agriculture,medicine,petrochemical,food and other fields.Because there is not necessarily a correlation between certain bands in the NIR and the material composition.And there are multiple correlations between various wavelength points.As a result,selecting a specific wavelength point or wavelength interval for modeling helps to improve the performance of the model.This article addresses the shortcomings of existing wavelength selection algorithms.Combining neural network and dictionary learning,two new methods of wavelength selection are proposed.The algorithm was verified by corn and oil dataset.And two wavelength selection methods are used in oil NIR analysis software.The main contents are as follows:(1)Some existing wavelength selection methods are studied,and some problems existing in them are analyzed.At the same time,some other near-infrared spectroscopy methods are studied,including near-infrared spectroscopy principles,denoising and baseline correction preprocessing methods,and regression models.(2)Combining the attention mechanism and neural network,a new method of sparse wavelength selection for near-infrared spectroscopy is proposed——WSNet.The algorithm uses the attention mechanism to extract feature information,combined with neural network self-learning capabilities,and adapts to the wavelength points that contain feature information.It is verified by the near-infrared spectroscopy data set of corn and crude oil.The experimental results show that the regression model established by the wavelength points selected by the WSNet algorithm has better performance.Then compared with the traditional wavelength selection algorithm UVE,WSNet has achieved better results on both data sets.Further proves the effectiveness of WSNet.WSNet further proved effectiveness.(3)A wavelength selection algorithm based on multi-dictionary learning is proposed——MDWLS.The algorithm uses the evenly divided NIR wavelength range to build a dictionary.The contribution weight of each wavelength interval to the spectrum is obtained through the dictionary,and then the wavelength interval with high contribution to the spectrum is selected.The experimental results show that the regression model based on the subset of wavelength intervals obtained by the MDLWS algorithm has better predictive ability.And compared with WSNet and UVE,MDLWS and WSNet achieved better results on the two data respectively.It is verified that MDLWS is beneficial to the improvement of regression model performance.(4)In response to the demand in the process of oil extraction,the oil NIR software was designed and implemented.And it applies the two wavelength selection methods proposed in this article to the software.The oil NIR analysis software implements the commonly used pre-processing methods and regression models for near-infrared spectroscopy.Finally,the software is tested to verify its usability. |
