What is it about?
Glass insulators support the live conductors on high-voltage power lines and keep them isolated from the towers. When salt, dust and grime build up on their surface and the air is humid, they leak small currents and produce faint electrical discharges (the "corona effect"), which in severe cases can escalate into a flashover and cause an outage. That corona discharge glows in the ultraviolet (UV), and its appearance reflects how polluted the surface is. We trained a convolutional neural network (CNN) to look at UV images of this glow and automatically classify the pollution severity into three levels — very weak, weak, and moderate. The model learned from more than 82,000 UV image frames captured in controlled high-voltage laboratory tests on artificially polluted insulators (salt plus kaolin, following IEC standards for pollution testing). It reached 98.57% accuracy with a very low error. Notably, because the network only needs the shape and position of the corona glow, low-resolution greyscale images work fine, making it fast (about 7 ms per image) and light enough for real-time use.
Featured Image
Photo by Keagan Henman on Unsplash
Why is it important?
The approach is both practical and a little surprising. It is non-invasive: pollution is judged from the corona's UV signature, without contact or laboratory salt-deposit sampling, and without de-energizing the line. It needs no manual feature engineering and clearly outperformed the alternatives tested — traditional image-processing methods (85–93%) and handcrafted-feature machine learning (92–94%) — and even beat fine-tuned ResNet50 (95.83%) and GoogLeNet (96.92%) while being lighter and faster to train. Strikingly, UV corona images look abstract and patternless to the human eye, yet the CNN classifies them reliably — suggesting the discharge carries real, learnable information about surface pollution. Because the model is fast and lightweight, it is well suited to field-deployable monitoring, such as substation inspection modules or UAV-mounted devices, enabling continuous, automated assessment and earlier maintenance before pollution flashover occurs — a leading cause of unplanned outages.
Perspectives
What I find most compelling here is the meeting point between electrical instrumentation and applied AI. UV corona images are essentially noise to a human observer — there is no obvious pattern to grade — and yet a well-designed network extracts the pollution severity from them with high reliability. That gap between what we can see and what the model can learn is exactly what drew me to the problem. I see this as a foundation rather than an endpoint: the natural next steps are taking it out of the controlled chamber into real field conditions and weather, adding the heavier pollution classes defined by the IEC standards, and applying explainability tools such as Grad-CAM so we can understand which regions of the corona the network actually relies on — turning a black-box classifier into a trustworthy inspection aid.
Prof. Dr. Eduardo Costa da Silva
Pontificia Universidade Catolica do Rio de Janeiro
Read the Original
This page is a summary of: CNN Classifier for Pollution Level in Glass Insulators Operating at High Voltage Alternating Current, Electronics Letters, January 2025, the Institution of Engineering and Technology (the IET),
DOI: 10.1049/ell2.70327.
You can read the full text:
Contributors
The following have contributed to this page
