The surface pollution will affect the insulator's insulation ability and harm the power system. The current detection method requires power outage sampling, which is tedious and time-consuming. Hyperspectral analysis can realize non-contact non-power outage detection and has good potential for application in insulator pollution detection. In order to reduce the data processing amount of insulator surface pollution hyperspectral detection and improve the accuracy of insulator hyperspectral data classification and identification, according to the current manual test sample preparation standards, Three artificial pollution samples of insulators with salt density of 0.22 milligrams per square centimeter, ash density of 0.1 milligrams per square centimeter and salt density of 0.3 milligrams per square centimeter and ash density of 0.1 milligrams per square centimeter were made. Hyperspectral sampling was carried out on insulator samples of different pollution levels. 15 regions were selected on each sample to extract regional spectral data, and a total of 90 groups of spectral data were obtained. 63 training set samples and 27 test set samples were selected. Competitive adaptive reweighted sampling (CARS) Algorithm, Successive Projections Algorithm (SPA and Uninformation Variables Elimination (UVE) were used to screen the characteristic wavelengths of insulator hyperspectral data and build a support vector machine classification model. Multivariate scattering correction (MSC), standard normal variation (SNV), first order derivative, deconvolution, moving average filtering, baseline correction, normalization, and wavelet transform were used to preprocess the spectral data. The classification experiments were carried out using the tested sample data, and the classification effects of different pretreatment methods and feature wavelength screening methods were compared. The experimental results show that the pre-treated data can improve classification recognition accuracy, and the pre-treated data's recognition accuracy can reach a low of 51.85% and a high of 96%, higher than the 40.74% in the untreated condition. Feature screening can reduce the dimensionality of the original spectral data, and SPA is the most efficient of the three methods, with an average screening rate of 3.56%. After screening, the accuracy of classification recognition can be improved, and the accuracy of classification recognition of original data can be increased from 40.74% to 74.07% after CARS screening. Pretreatment of spectral data combined with feature wavelength screening can greatly improve the classification and identification accuracy of dirty insulators. The classification and identification accuracy of MSC-CARS, SNV-CARS, MSC-UVE, and normalized UVE can reach 100% after combined processing. The average classification accuracy of CARS, SPA, and UVE combined with 8 preprocessed test sets was 87.05%, 86.25%, and 83.47%, respectively, indicating that the combination of preprocessing and feature screening plays an important role in improving data quality and model performance. These preprocessing and characteristic wavelength screening methods of spectral data can effectively reduce the original data dimension and simplify the model complexity, which can play an important role in improving the accuracy of insulator pollution classification and recognition.