There are problems of large average error and low classification accuracy in feature space extraction of transformer oil-paper insulation comprehensive diagnosis. These problems are due to the existence of correlation and redundant features in high-dimensional feature space. Thus a feature quantity optimization strategy based on a fast filtering correlation algorithm and limit gradient rise is proposed. First, from the measured data of transformer dielectric response, various kinds of time-domain dielectric characteristics are extracted to form the initial high-dimensional feature space. Secondly, a two-stage time-domain feature selection method is proposed. In the first stage, a fast correlation filtering algorithm is used to eliminate the features with low correlation and high redundancy, and in the second stage the importance of features is evaluated according to the limit gradient, so as to determine the optimal feature space. Finally, different control groups are set for comparative demonstration of the optimal feature space. This effectively verifies the rationality and accuracy of the optimal feature space obtained by adopting the optimal strategy proposed above. This work is supported by the National Natural Science Foundation of China (No. 61174117).