An Approach to Condition Assessment of High-Voltage Insulators by Ultrasound and an Ensemble of Convolutional Neural Networks

This paper proposes an approach and proof of concept for evaluating the condition of high-voltage insulators of power distribution networks (up to 145 kV) using ultrasonic tests provided by a probe equipment in a methodology based on an ensemble of convolutional neural networks and robust preprocessing techniques. It presents the laboratory tests and the conditions in which several real situations were simulated. Next, pre-processing, the neural network architectures and the flowchart of the insulation condition analysis methodology are detailed. Finally, the results of the diagnostics from the methodology with the training, validation and test sets are presented and discussed. The proposed methodology achieved 100 % accuracy in validation and test data.