The key gas method and IEC method for dissolved gases analysis only focus several gas index that may be affected by random error and relative percentages, and maybe it is not perfect to be used to conduct preventive maintenance and inspection of transformers. To solve the problem, key gases are mapped into 62 features including traditional combinations of key gases and other new features. The features highly correlated to transformer status are selected by maximum information coefficient, to escape random error of gas concentration. To avoid redundancy between selected features, Pearson coefficient is used to filter features to reduce the redundancy of the selected features. The result showed that correlation coefficient between the selected features and the transformer fault is relatively high. And the result of distance correlation coefficient also shows that the feature set formed by the selected features is more closely related to the transformer overheat fault state than that of the traditional fault characteristic gas.