Raman Spectral Recognition Technology Based On Improved Nonlinear Partial Least Squares Method

Raman spectroscopy is a scattering spectrum with the ability to characterize molecular structures.Substances can be identified by processing Raman spectral data.Partial least squares(PLS)is one of the most widely used methods.In practical applications,due to the interference of nonlinear factors such as nonlinear noise and substance concentration,the traditional linear method will no longer be applicable.To solve this problem,the nonlinear partial least squares method came into being.However,the current existing nonlinear methods mainly focus on improving the nonlinear processing capability and recognition effect,and the robustness and computational complexity of the algorithm have not been fully considered.Aiming at the above problems,this thesis proposes a novel graph-based nonlinear partial least squares discriminant method,that is,graph partial least squares(GPLS).And the validity of the method in this thesis is verified by building a longdistance pulsed Raman detection system independently and collecting spectral data.The specific content is divided into the following parts:Inspired by the laplacian eigenmaps method,GPLS improves PLS by combining manifold learning and graph theory.GPLS establishes the neighbor relationship through the label information,draws the principal components of the independent variable and the dependent variable closer to each other,and re-establishes the objective function.Through the derivation of the objective function,the relationship between GPLS and traditional PLS is deeply discussed in theory.Traditional PLS is a special case of GPLS,and the two have been unified in the field of graph theory.And under the inspiration of traditional PLS,GPLS has carried out its generalization under the condition of non-simple graph.In addition,GPLS is still highly efficient,and its computational complexity is comparable to traditional linear methods.The thesis establishes a graph-based Raman spectral recognition framework(GRSR)with GPLS as the core.In this framework,in order to effectively reduce the interference of non-systematic errors and enhance the robustness of signal smooth representation,this thesis proposes an iterative graph filtering algorithm.Firstly,by extending the existing adaptive neighbor graph method under supervised conditions,the label information and the distance relationship between signals are effectively fused,and then the iterative graph filtering algorithm is obtained by combining traditional graph filtering.Through the filtering algorithm,the smooth representation of the training data is firstly obtained,and the smooth representation is used for GPLS modeling.Afterwards,the smooth representation and the test data are unified to build a graph,and the idea of K-Nearest Neighbor algorithm is combined with the connection relationship in the graph to optimize the feature representation of the test data,thereby reducing the burden of downstream recognition tasks.In order to verify the effectiveness of the method in this thesis,simulation verification was first carried out on a synthetic dataset.In addition,by independently building an ICCD-based remote pulsed laser Raman detection system,the Raman spectrum data of various minerals were successfully collected in an outdoor environment,and the detection distance reached 100 m.Afterwards,extensive comparative experiments were designed and carried out on mineral Raman spectral data and public Raman spectral datasets under different signal-to-noise ratios.The experimental results show that the method in this thesis is not only superior to other nonlinear PLS methods in terms of recognition effect and robustness,but also has high efficiency.