Feature Selection and Classification Optimization of Transformer Oil Odor Data With Recursive Feature Elimination Using Grid Search and Cross-Validation

Power transformers are vital for transmitting, distributing, and using energy produced in electrical energy systems. The failure of a distribution transformer can cause significant financial losses to society and cause irreparable problems. Therefore, power transformers' reliability, continuous monitoring, and fault-free operation are critical. In this study, an electronic nose system was developed to detect the duration of oil usage in power transformers based on its smell. In the system designed with eleven inexpensive gas sensors, a total of 200 transformer oil odors with four different usage periods were analyzed. Four features were selected from eighty-eight features using the Recursive Feature Elimination method with Grid Search and Cross-Validation, and they were classified with six different classifiers. With the Extra Trees algorithm, the most successful classifier, classification performance of 0.9810 CA, 0.9810 SE, 0.9937 SF was achieved without feature selection, and 0.9610 CA, 0.9610 SE, 0.9870 SF was achieved by selecting features. Dielectric breakdown voltage tests of the oils in the study were also performed, and the results supported the results of the electronic nose system. According to the results obtained, it can be concluded that transformer oil maintenance can be performed economically, practically, and reliably with the proposed system.