Study On Multi Scale Feature Extraction And Integrated Enhanced Neural Network Diagnosis Of Oil-Paper Insulation Aging By Raman Spectroscopy

Oil-paper insulation power equipment is one of the core of power system.Its life mainly depends on the aging state of oil-paper insulation system.Accurate and effective aging diagnosis is of great significance to ensure the safe and stable operation of power system.There are many information about oil-paper insulation aging in Raman spectra,and the Raman spectral features of oil-paper insulation aging are obvious.However,the corresponding relationship between Raman information and oil-paper insulation aging needs to be further explored,and the oil-paper insulation aging diagnosis method based on Raman spectroscopy is not perfect.In this paper,the aging mechanism of oil-paper insulation and its Raman correlation were studied.The data preprocessing method of the Raman spectra of oil-paper insulation aging was studied,and a balanced oil-paper insulation aging Raman spectral database has been established.Under the supervision of aging degree,the Raman features of oil-paper insulation aging were extracted from multiple scales.An Adaboost BP diagnosis model of oil-paper insulation aging was established,and the diagnosis correction model of "old paper new oil" in different states was established based on generalized regression neural network.The main work of this paper is as follows:1.Based on molecular dynamics simulation,the aging process of oil-paper insulation was analyzed and Raman correlation was studied.Based on Reax FF force field,the oil-paper insulation model(11 molecules)including cellulose,alkanes,cycloalkanes,aromatics,etc.was constructed.Chemical reactions that may occur in the aging process of oil-paper insulation were simulated,and the characteristics of the main products in the aging process were analyzed.Their Raman activities were calculated.The formation rules of Raman peaks caused by vibrations of C-H bond,C-C bond,C=C bond,C-O bond and C=O bond were obtained.2.Pretreatment of Raman spectral data and establishment of Raman database of oil-paper insulation aging.According to the characteristics of oil-paper insulation Raman spectral data,the methods of peak recognition based on derivative spectrum,peak removal based on cubic curve,smoothing based on median value of three-point cyclic fast Fourier transform and spectrum normalization based on maximum and minimum value were proposed.Based on the processed Raman spectral data,the adaptive sample synthesis method was studied to improve the distribution of samples in each aging stage in the oil-paper insulation aging Raman spectral database.A balanced oil-paper insulation aging Raman spectral database covering the whole life cycle of oil-paper insulation was established.3.Multi scale Raman feature extraction and analysis of aging of oil-paper insulation.Under the supervision of the aging degree labels of oil-paper insulation samples,the random forest method for selecting the original Raman spectral features of aging of oil-paper insulation,the linear discriminant analysis method for extracting the Raman category features of oil-paper insulation at different aging stages,the partial least squares algorithm and quadratic mutual information method for extracting the features that mapping the degree of polymerization of oil-paper insulation were studied.The redundant information in the Raman spectra of oil-paper insulation was eliminated,and the strong correlation between the extracted features and the aging of oil-paper insulation has been retained.The intrinsic relationship between the aging features of oil-paper insulation and the aging process was analyzed.4.Establishment and modification of Raman spectroscopy integrated enhanced neural network diagnostic model for oil-paper insulation aging.Based on the known aging state of the training samples and the extracted Raman features,a single aging diagnostic model based on BP neural network with additional momentum was constructed.By using the serial adaptive integrated enhanced algorithm and combining multiple weak learners through the sample weight transformation guided by the prediction effect,the corresponding relationship between the Raman features and the continuous labels of oil-paper insulation aging degree is fully mined,and the Raman adaptive integrated enhanced back propagation neural network oil-paper insulation aging diagnosis model is established.For the "old paper and new oil" system at different states after oil replacement,the diffusion interaction of aging features in the old paper and the fresh oil was studied.The influence of different states on the previous diagnosis results was studied.A modified model was constructed by using generalized regression neural network.The correction of the diagnosis of the "old paper and new oil" samples in different states was realized.To sum up,the formation rules of Raman peaks caused by the main chemical bond vibration throughout the aging process were obtained,a balanced oil-paper insulation aging Raman spectral database has been established,the physical significance of Raman spectral features in the process of oil-paper insulation aging was explored,and a Raman diagnostic model of oil-paper insulation aging was established.The research of this paper provides a new idea for the accurate aging diagnosis of oil-paper insulation,which is conducive to the on-site or even on-line aging diagnosis of oil-paper insulation power equipment.The excellent performance of oil-paper insulation aging diagnosis based on Raman spectroscopy provides a strong support for the life cycle management of oil-paper insulation power equipment.