This study was aimed to improve the faults diagnosis capability of the dissolved gas analysis (DGA) system installed for the oil-filled transformers used in Formosa Plastics Corporation. Three attempts were made: the modification of Laborelec ratio codes criterion for achieving effective and accurate DGA and faults diagnosis, the adoption of rough set theory (RST) for improving knowledge mining capacity and the addition of feedback mechanism for enhancing the overall capacity of the entire system. Main achievements of this study included: (1) improving the crisp classification accuracy from 45.1% to 100%; (2) improving the capability of extracting and discovering experts' experience up to 90.7% (with CF=1 and γ (subscript p) =90.8%).