Pixel-Level Assessment Model of Contamination Conditions of Composite Insulators Based on Hyperspectral Imaging Technology and a Semi-Supervised Ladder Network

Real-time and accurate supervision of contamination conditions of composite insulators can effectively prevent the occurrence of pollution flashover accidents and ensure stable operation of the power system. In this article, a refined and controllable artificial contamination experiment is carried out, in which a total of 60 types of pollutants (12 contamination conditions) are formed by combining 20 types of equivalent salt deposit densities (ESDDs) at four contamination levels and three types nonsoluble deposit densities (NSDDs)/ESDD ratios to pollute composite insulators. Hyperspectral imaging technology is also introduced to extract the spectral data of polluted insulators, and the changing pattern of spectral curves under different pollution conditions is studied. Then, the spectral curve is preprocessed by black-and-white correction, standard normal variate (SNV), and detrending correction, and the preprocessed data are analyzed for band correlation. According to the variation pattern of the spectral curve and the band correlation characteristics, the full set of bands is divided into red, other visible, and near-infrared regions for principal component transformation respectively. Compared with the traditional principal component transformation, the former can reduce the calculation amount by 66%. Finally, to solve the problem of insufficiently labeled samples in practical engineering, a semi-supervised ladder network is used to realize pixel-level evaluation of insulator contamination conditions, and evaluation results are visualized. Research results show the pixel-level evaluation method can accurately evaluate contamination conditions of different areas on the surface of composite insulators. A semi-supervised model can guarantee the accuracy of classification results when there are few labeled samples. When the supervision rate is 10%, the classification accuracy of a semi-supervised model for contamination grade, NSDD/ESDD ratio, and contamination condition is higher than 95%, while that of a supervised model is lower than 90%.