Adaptive neuro-fuzzy inference system approach for simultaneous diagnosis of the type and location of faults in power transformers

Electrical, mechanical, and thermal stresses can degrade the quality of the insulation in power transformers, causing faults [1]. Several methods are used for fault diagnosis in transformers, e.g., dissolved gas analysis (DGA), measurement of breakdown voltage, and tan δ, pollution, sludge, and interfacial tension tests [2]. Of these, DGA is the most frequently used. Thermal and electrical stresses result in fracture of the insulating materials and the release of several gases. Analysis of these gases may provide information on the type of fault. Various standards have been suggested for the identification of transformer faults based on the ratio of dissolved gases in the transformer oil, e.g., International Electrotechnical Commission (IEC) standards [3]–[7], and these standards has been quoted in many papers, e.g., [8]–[15]. However, they are incomplete in the sense that, in some cases, the fault cannot be diagnosed or located accurately. Intelligent algorithms, e.g., wavelet networks [16], neuro-fuzzy networks [17], [18], fuzzy logic [8], [12], and artificial neural networks (ANN) [2], [9], [10], [19], [20] have been used to improve the reliability of the diagnosis. In these algorithms, the type of fault is diagnosed first, and the fault is then located using the ratio of the concentrations of CO2 and CO dissolved in the transformer oil [21], [22]. The algorithms are not entirely satisfactory. The wavelet network has high efficiency but low convergence, the fuzzy logic method has a limited number of inputs and, in some cases, it is very difficult to derive the logic rules, and the ANN need reliable training patterns to improve their fault diagnosis performance. In this paper, we present a new method for simultaneous diagnosis of fault type and fault location. It uses an adaptive neuro- fuzzy inference system (ANFIS) [23]–[27], based on DGA. The ANFIS is first “trained” in accordance with IEC 599 [3], so that it acquires some fault determination ability. The CO2/CO ratios are then considered additional input data, enabling simultaneous diagnosis of the type and location of the fault. The results obtained by applying it to six transformers are presented and compared with the corresponding results obtained using ANN and some other standards and methods.