Modified Successive Geometric Segmentation Method Applied to Power Transformers Faults Diagnosis

ABSTRACT This paper proposes a new approach to address the complex problem of the internal faults classification in power transformers. The fault classification is made using an algorithm based on a geometric approach to construct an artificial neural network. The input parameters are related to the concentrations of various gases dissolved in insulating oil. The proposed technique is able to generate both the topology and the weight for each neuron without specifying initial network parameters. Datasets of dissolved gas concentration in high-voltage transformers that have been found in the scientific literature are used to validate the methodology. The results indicate a high accuracy rate compared to other important computational intelligence methods used to high-voltage transformers faults classification.