Induction of Decision Trees to Diagnose Incipient Faults in Power Transformers

This article describes the use of a decision tree (DT) approach based on computational intelligence (CI) for the analysis and diagnosis of incipient faults in power transformers. The method relies on measuring the combustible gas concentration in parts per million (ppm) from samples of the insulating oil. Currently, the Duval triangle method is one of the most traditional techniques used for the dissolved gas analysis (DGA), but it has limited accuracy. To overcome such conventional performance problems, CI techniques as artificial neural networks, fuzzy systems, and more recently DTs have been proposed as possible solutions. In this context, by using the gain ratio as a metric for attribute selection, this work demonstrates that the DT algorithm is capable of extracting as much information as possible from each class. It can also provide a solution for unsolved cases by using the traditional diagnosis method. Overall, it is reasonable to state that DT plays an important role in the improvement of DGA performance, this being a promising tool that can be combined with other traditional techniques. Another important aspect is that in the end of training the tree generates clear and ease-of-use rules. The proposed technique results in fast and accurate solutions for diagnosing faults in power transformers. It can be effectively implemented in corrective maintenance to avoid permanent burning and damage of equipment.