Monitoring ZnO surge arresters using convolutional neural networks and image processing techniques combined with signal alignment
Bruno Vinícius Silveira Araújo, Gustavo Aragão Rodrigues, Johnny Herbert Paixão de Oliveira, George Victor Rocha Xavier, Ulisses Daniel Enes de Souza Lebre, Charles Antony Cordeiro, Eduardo Oliveira Freire, Estácio Tavares Wanderley Neto, Tarso Vilela Ferreira · Measurement · 2025
• Consistent detection and segmentation of thermal anomalies in ZnO Surge Arresters. • Method robust to environmental influences and fluctuations in distance and angle. • Identifies defects through temperature analysis, predicts evolution, enables alerts. • Enhances maintenance and safety, identifies anomalies in various high-voltage gear. • Thermal Profile Alignment for Improved Monitoring of Electrical Equipment. Monitoring and preemptive maintenance of high-voltage electrical equipment play a crucial role in preventing breakdowns and ensuring the seamless operation of power systems. Infrared monitoring stands out due to its convenience, safety protocols, and utilization of temperature as a key indicator for assessing the structural integrity and component health of equipment. This paper introduces a method for monitoring ZnO lightning arresters by analyzing their thermal profiles. The approach involves employing a convolutional neural network and computer vision processes to detect, segment, and extract thermal data from these devices. An alignment algorithm facilitates comparison and classification of operational integrity. The method provides an accurate and comprehensive evaluation of ZnO lightning arresters, enabling continuous monitoring and efficient diagnosis. By analyzing thermal imaging data from a 500 kV substation and conducting laboratory tests on both healthy and intentionally defective equipment, the detection algorithm achieves favorable precision rates (0.861), recall (0.855), mAP50 (0.903), and mAP50:95 (0.615), ensuring precise detection and segmentation. Despite variations caused by thermal imager measurement errors, distance fluctuations, angle deviations, and environmental factors, the algorithm consistently identifies both normally operating and faulty equipment. Notably, when assessing healthy and intentionally defective equipment, the method achieves excellent accuracy in anomalies and its locations.