Quantitative Method Of Back Propagation Artificial Neural Network For Transformer Fault Characteristic Gas Raman Spectroscopy Detection

The content and components of transformer fault characteristic gases（CO₂,CO,H₂,CH₄,C₂H₆,C₂H₄,C₂H₂）are important indicators for diagnosing power transformer faults and aging,and the accurate detection of transformer fault characteristic gas content is important for analyzing early latent faults and ensuring safe transformer operation.Raman spectroscopy gas detection has shown good application prospects,but accurate quantitative methods lack in-depth research.The existing spectral pre-processing process relies on manual selection of parameters,which lacks consistency and accuracy in mass processing;Raman spectral quantification does not consider the influence of peak intensity including laser power and temperature,and mostly establishes a linear relationship between peak intensity and concentration,and the accuracy needs to be improved.The paper carries out the research on the quantitative method of Raman spectroscopy detection of transformer fault characteristic gases based on back propagation artificial neural network,the main contents are as follows.1.To study the baseline correction method based on sparsity with baseline modeling as the low communication number and the spectral peak fitting method based on Gaussian fitting,to establish a Raman spectral peak intensity calculation method with self-selection of key parameters,and to effectively reduce the influence of human factors on spectral pre-processing.2.The Raman peaks of seven transformer fault gases（CO₂,CO,H₂,CH₄,C₂H₆,C₂H₄,C₂H₂）were selected based on a hollow-core anti-resonant fiber-enhanced Raman spectroscopy platform,and the characteristics of gas partial pressure,laser power and temperature were studied.The Raman peak intensity（peak height and peak area）of each gas showed a good linear relationship with the gas partial pressure and laser power;the peak intensity and frequency shift of each gas in the experimental temperature range（20℃to 90℃）decreased approximately linearly with the increase of temperature;a sample library of peak intensity of Raman spectra of transformer fault characteristic gases was formed.3.A quantitative analysis model of Raman spectra of transformer fault characteristic gases based on back propagation artificial neural network（BP-ANN）is established;by comparing different optimization functions,loss functions and network connections,a combination of optimization function as Adamax,loss function as Smooth L1loss,2 hidden layers and 50 and 100 neurons,respectively,is selected for seven transformer The errors of the fault characteristic gases under the quantitative analysis model are analyzed,and it is found that the min errors of the test set without Z-Score normalization are 0.3,1.25,0.77,10.4,0.14,0.09,and 0.58,respectively;the min errors of the test set after Z-Score normalization are 4.69×10^-4,1.44×10^-5,5.29×10^-4,1.92×10^-5,1.81×10^-3,5.35×10^-4,1.50×10^-3;the quantitative accuracy of each of the seven transformer fault characteristic gases averaged 91.1%at 100 Pa partial pressure and 95.4%at 500 Pa partial pressure.The quantitative analysis method of Raman spectral back propagation artificial neural network studied in the paper lays the foundation for accurate detection of transformer fault characteristic gas Raman spectra.