Extracting feature parameters from the recovery voltage polarization spectrum is currently a widely adopted method for evaluating the status of transformer oil-paper insulation. However, the polarization spectrum is prone to anomalous feature data due to factors such as working condition interference and artificial errors, which seriously reduces the accuracy of the evaluation. In response to the above issues, this paper proposed a recovery voltage data cleaning method based on local outlier factor (LOF) and improved K-nearest neighbor (IKNN). Firstly, Maximum recovery voltage Urmax, the initial slope Sr and dominant time constant tcdom of the recovery voltage polarization spectrum were selected as aging feature parameters, and anomalous feature data in the non-standard polarization spectrum were identified and filtered out based on the LOF algorithm. Secondly, the Fuzzy C-means (FCM) clustering algorithm was used to reduce the interference of noise points on the KNN algorithm, and the correlations between various features were highlighted by weighted Euclidean distance scale. Then, a data filling model architecture based on IKNN was constructed to fill in missing feature data. Finally, multiple sets of measured data were incorporated to validate the effectiveness of the proposed data cleaning method. The results indicate that the accuracy of status evaluation after data cleaning has increased by about 50% compared to the original data, which effectively improves the quality of transformer recovery voltage data and lays a solid foundation for accurate perception of transformer operation status.