Abstract:
High voltage power distribution lines play a fundamental role in electricity transmission.
They are quite essential to transmit the electrical power from the generation units to the various
distribution stations, and their proper functioning being vital for the operation of power grids.
High voltage insulators are the essential components that serve as electrical insulation and
mechanical support for live conductors. Since the insulator surfaces are physically unprotected
and regularly exposed to severe weather conditions, their monitoring and maintenance become
a fundamental concern for the power transmission companies.
In this thesis, a comprehensive literature review over previously established industrial
methods in the monitoring of outdoor insulators is conducted to address this problem. Then,
state-of-the-art and current research methods along with advanced intelligent techniques are
discussed. A new methodology to evaluate the outdoor insulation surface is introduced and
discussed, and in-depth analysis using different classifiers is conducted over data obtained from
ceramic, glass, and polymer insulators in various weather conditions. A dataset composed of
3000 high voltage outdoor insulator images with different surface conditions such as clean
surface, contamination with soil, cement, and water is created to construct an insulator surface
classifier model. Utilizing recent techniques in deep learning, this thesis proposes the following
methods: brute-force model selection, pre-trained CNN, and model optimization steps applied
to the best model selected from the brute-force model selection technique. Finally, keeping in
mind the possible implementation of the best selected CNN in resource-limited embedded
devices, a model complexity reduction technique is applied to decrease the storage and
computation process by training only the Dense classifier layers. Ultimately, the technique
employed in this thesis can distinguish a normal insulator from a contaminated insulator and
advise the type of pollution.
The thesis results show that the brute-force model selection method can adequately
construct a highly efficient solution with 90.2% accuracy on the unseen test dataset. Also, the
optimization technique employment led to three times lighter architecture at the expense of
~4.9% reduction in the accuracy.