Abstract
The representative dissolved gases analysis (DGA) method for transformer fault detection faces many shortcomings in early fault diagnosis, which restricts the application and development of fault detection technology in the field of transformers. In order to diagnose early failure in time, fluorescence analysis technology has recently been used for the research of transformer failure diagnosis, which makes up for the shortcomings of DGA. However, most of the existing fluorescence analyses of insulating oil studies combined with intelligent algorithms are a qualitative diagnosis of fault types; the quantitative fault diagnosis of the same oil sample has not been reported. In this study, a typical fault simulation experiment of the interval discharge of insulating oil was carried out with the new Xinjiang Karamay oil, and the fluorescence spectroscopy data of insulating oil under different discharge durations were collected. In order to eliminate the influence of noise factors on the spectral analysis and boost the accuracy of the diagnosis, a variety of spectral preprocessing algorithms, such as Savitzky-Golay (SG), moving median, moving mean, gaussian, locally weighted linear regression smoothing (Lowess), locally weighted quadratic regression smoothing (Loess), and robust (RLowess) and (Rloess), are used to smooth denoise the collected spectral data. Then, the dimensionality reduction techniques of principal component analysis (PCA), kernel principal component analysis (KPCA), and multi-dimensional scale (MDS) are used for further processing. Based on various preprocessed and dimensionally reduced data, transformer failure diagnosis models based on the particle swarm optimization algorithm (PSO) and the secretary bird optimization algorithm (SBOA) optimized BPNN are established to quantitatively analyze the state of insulating oil and predict the durations of transformer failure. By using the mathematical evaluation methods to comprehensively evaluate and compare the effects of various algorithm models, it was found that the Loess-MDS-SBOA-BP model has the best performance, with its determination coefficient (R(2)) increasing to 99.711%, the root mean square error (RMSE) being only 0.27144, and the other evaluation indicators also being optimal. The experimental results show that the failure diagnosis model finally proposed in this paper can perform an accurate diagnosis of the failure time; the predicted time is closest to the true value, which lays a foundation for the further development of the field of transformer failure diagnosis.