A CORONA RECOGNITION METHOD BASED ON VISIBLE LIGHT COLOUR AND MACHINE LEARNING

Abstract

Can we detect electric discharge states in gases based on the information on visual images? The
document deals with a new kind of method where we build several detection models for different states of corona
discharge by applying four kinds of machine learning algorithms to extract color, brightness, and shape
information characteristics of visible images taken by a digital camera. Every model is then tested on a new set of
images to measure its performance. The four different machine learning algorithms are Support Vector Machine
(SVM), K-Nearest Neighbor regression (KNN), Single Layer Perceptron (SLP), and Decision Tree (DT)
algorithms.
In the existing system the traditional research of the discharge image was conducted from a
qualitative point of view, such as morphology description, strong or weak light intensity, etc. However, only a
few studies focused on quantitative evaluation can be found. With the development of computer techniques, the
digital image processing methods have been applied extensively to study discharge characteristics, such as
breakdown paths, discharge area, etc., especially in ultraviolet (UV), which can help tackle complex problems by
using statistical techniques or the fractal theory. Although the images obtained by a high-speed camera
(nanosecond time scale) can provide some details of a single discharge, the essence of gas discharge remains
random under the same macroscopic physical conditions.
The proposed system proposes a method where we build several detection models of different states
of corona discharge by applying machine learning algorithms to extract the color, brightness, and shape
characteristics of visual images. In the second part, the experimental set of corona discharge is introduced. In the
third part, the idea of three primary color- [red, green, and blue (RGB)] gray level histogram (RGB-GLH) of
visual images will be introduced and the specific process of applying machine- learning algorithms to analyze
the characteristic information of visual images will be discussed. In the fourth part, the prediction results of our
model are reported and compared.