A feature parameters extraction method of Partial Discharge (PD) Ultra-High Frequency (UHF) signal based on Dual?Tree Complex Wavelet Transform (DTCWT) and Locally Linear Embedding (LLE) is proposed. The typical transformer internal insulation defects can be recognized effectively based on the characteristic parameter. At first, the PD UHF signals of transformer are decomposed by DTCWT, and a series of detail component signals at different scales of variation can be obtained. The skewers and kurtosis of every detail component signal are extracted as the original feature parameters. Through the LLE method, the eigenvector consist of original feature parameters can be dealt with dimension reduction to get the final feature parameters and eigenvector. The eigenvector are inputted into the support vector machine to pattern recognition. The results show that the propos feature parameters can be used to recognize the typical PD UHF signal in transformer, the overall recognition accuracy of simulated insulation defects reached 98.35%, and the overall recognition accuracy of field test signals reached 92.1%. This work is supported by Science and Technology Project of State Grid Corporation of China (No. 2018A-058).