Data quality is an important factor affecting the accuracy of transformer fault diagnosis. In order to solve the problem of missing on-line monitoring data for transformer dissolved gas, an on-line fault diagnosis method using loop iterations of improved k-nearest neighbors and multi-class SVMs for power transformer based on missing data repair is proposed. In the k-nearest neighbor method, the Manhattan distance which is weighted by the negative exponent of the correlation coefficient is used to measure the distance between samples. On one hand, the influence of the strong correlation indexes on the missing data can be highlighted to improve the accuracy of data repair. On the other hand, the improved Manhattan distance is suitable for an efficient search strategy based on k-d tree, which can achieve fast search for massive historical data and meet the real-time demand of on-line diagnosis. Diagnosis test results show that the proposed method can reduce the influence of missing data on the accuracy of transformer fault diagnosis and realize the accurate and efficient on-line diagnosis for transformer fault. This work is supported by National Natural Science Foundation of China (No. 51807173) and Mass Innovation Project of State Grid Zhejiang Electric Power Company Ltd. (No. 5211HZ18004L).