Use of decision tree algorithms to diagnose incipient faults in power transformers

This paper describes the use of a decision tree based on Computational Intelligence methodology for the analysis and diagnosis of incipient failures in power transformers by using the concentrations in ppm of the combustible gases present in samples of transformer oils. It is well known that power transformers are one of the most, technically and economically, relevant equipment in a transmission and distribution electric plant. It is essential to ensure the electric plant continuous operation and prevent possible failures that may occur because of their natural life cycle or electrical arrangement that are submitted. A great effort should be done to avoid this equipment outage. Currently, Duval Triangle method is one of the most used traditional techniques for Dissolved Gas Analysis, however this technique has shown limited accuracy. To overcome the conventional performance problems, Computational Intelligence (CI) techniques as neural networks, fuzzy systems and more recently Decision Trees (DT) have been proposed as methods for DGA analysis. This work have shown that the DT algorithm, by using the gain rate as a metric for attribute selection, have been able to extract as much information as possible from each class, and that the algorithm can provide a solution for unsolved cases by using traditional diagnose method. Finally, it can be concluded that the DT technique plays an important role for improving DGA analysis and it appears to be a promising tool by itself and in conjunction with traditional techniques. Another important aspect of using DT is that in the end of training the tree generates clear and ease-of-use rules for DGA diagnosis. The proposed DT approach results in a fast and accurate technique for diagnosing power transformer failure that can be effectively implemented in corrective maintenance to avoid permanent burning and equipment destruction.