SEGMENTACIJA IZOLATORA NA TERMALNIM SLIKAMA UPOTREBOM METODA MAŠINSKOG UČENJA

Abstract

The infrastructure of power grids, which cover vast territories including
inaccessible areas, requires regular maintenance to ensure their proper functioning,
safety, and continuous power supply. Traditional inspection methods require the
presence of teams on-site, they are costly and dangerous. Therefore, automated
inspection is being used increasingly, including the use of unmanned aerial vehicles
(UAVs) that capture images of power lines from hard-to-reach angles and different
spectral ranges (visible and thermal). The obtained images are then analyzed
using machine learning algorithms that detect defects or faults in power line components.
The goal of using artificial intelligence is to improve and optimize inspections,
thereby increase safety, reduce costs, and enhance the efficiency of power
grid maintenance. Advanced machine learning techniques are used for the precise
detection of equipment parts (insulators, conductors, connectors, and suspension
equipment) and defects on them, which is achieved by analyzing RGB and thermal
images obtained by UAVs. This paper describes the segmentation of insulators in
thermal images using the PCNN (Pulse-Coupled Neural Network) algorithm. The
segmentation is first performed on bounding boxes detected by the YOLOv8 architecture,
adjusting the parameters of the PCNN algorithm and by further post-processing
that minimizes the background noise. The algorithm is then extended so
that it takes as input the thermal image, the coordinates of the region of interest
for the insulator, threshold values and parameters. The result is a generated binary
mask of the thermal image’s dimensions on which the insulator is segmented. The
detected insulator can then be used as part of a broader thermal analysis, which
monitors the temperature of the equipment so that, in case of irregularities an alarm
is triggered.