【正文】 rove the search efficiency by properly arranging the location of significant rules in the inference procedures. The latter strategy only searches the key conditional statements in the antecedent that are responsible for establishing whether the entire rule is true or false. Taking the advantages of these two approaches in the building and structuring of a knowledge base improves inference efficiency significantly. As for manmachine interface. KES has an effective interface which is better than typical knowledge programming languages, such as, PROLOG or LISP. With the help of this interface, the capability of tracing, explaining and training in an expert system is greatly simplified. 4. IMPLEMENTATION OF THE PROPOSED EXPERT SYSTEM An expert system is developed based on the proposed interpretative rules and diagnostic procedures of the overall system. To demonstrate the feasibility of this expert system in diagnosis, the gas data supported by MTL of TPC have been tested. In Taiwan, the MTL of TPC performs the DGA and sends the results to all acting divisions relating to power transformers. In return, these acting divisions are requested to collect and supply their transformer oil samples periodically. After analyzing oil samples, more than ten years39。 gas characteristic data are then used to judge the transformers39。 For example large amounts of C H and H are produced with minor arcing fault 4. quantities of CH 2aid C2H4 2 may be a symptom of an arcing fault. 3. THE PROPOSED DIAGNOSTIC EXPERT SYSTEM This study is aimed at developing a rulebased expert system to perform transformer diagnosis similar to a human expert. The details of system processing are described below. The Proposed Diagnostic Method Diagnosis is a task that requires experience. It is unwise to determine an approach from only a few investigations. Therefore, this study uses the synthetic expertise method with the experienced procedure to assist the popular gas ratio method and plete practical performance. Experienced Diagnostic Procedure The overall procedure of routine maintenance for transformers is listed. The core of this procedure is based on the implementation of the DGA technique. The gas ratio method is the significant knowledge source. Some operational limitations of the gas ratio method exist. The ratio table is unable to cover all possible cases. Minimum levels of gases must be present. The solid insulation involving CO and CO are handled separately and the gas ratio codes have been developed mainly from a freebreathing transformer. Other diagnostic expertise should be used to assist this method. Norms, synthetic expertise method and data base records have been incorporated to plete these limitations. The first step of this diagnostic procedure begins by asking DGA for an oil sample to be tested. More important relevant information about the transformer39。 hydrogen(H2), methane(C2H 2), ethane(C2H6), ethylene(C2H4), acetylene(C2H2) and carbon monoxide(C0). Many interpretative methods based on DGA to diagnose the nature of incipient deterioration have been reported. Even under normal transformer operational conditions, some of these gases may be formed inside. Thus, it is necessary to build concentration norms from a sufficiently large sampling to assess the statistics. TPC investigated gas data from power transformers to construct its criteria. The developed knowledge base in this paper is partially based on these data. On the other hand, Dornerburg developed a method to judge different faults by rating pairs of concentrations of gases, ., CH /H , C H /C2H4, with approximately equal solubility and fusion coefficients. Rogers established mare prehensive ratio codes to interpret the thermal fault types with theoretical thermodynamic assessments. This gas ratio method was promising because it eliminated the effect of oil volume and simplified the choice of units. Moreover, it systematically classified the diagnosis expertise in a table form. Table 1 displays the ratio method as proposed by Rogers . The dissolved gas may vary with the nature and severity of different faults. By analyzing the energy density of faults, it39。