To improve the diagnosis accuracy of power transformer, this work proposes a multi-level fault diagnostic technique that uses diffusion maps and extreme gradient boosting (XGBoost). Firstly, 14-dimensional DGA features, which are based on the fusion of the dissolved gas content (C2H6, C2H4, C2H2, CH4 and H2,) and 9 features constructed by non-code ratio method, are extracted. Then, a diagnostic model is established which is comprised of four levels, the first is for identifying faults or normal states, the second is for diagnosing discharge or thermal faults, the third is for classifying discharge faults, and the fourth is for recognizing thermal faults. There are two modules in each level model structure: feature extraction and fault diagnosis module. Feature extraction module adopts diffusion maps to determine the dimension redundancy. Fault diagnosis module built with XGBoost uses those features extracted from the feature extraction module as inputs. Finally, IEC TC 10 database and IEEE dataset are used to test the proposed multi-level diagnostic method, and the results demonstrate that it is more accurate at diagnosing faults than that of the other methods such as k-nearest neighbor, extreme learning machine, and probabilistic neural network.