In order to solve the problems of low resolution

weak anti-noise performance

and inaccurate fault location in discharge diagnosis of ultraviolet imaging

an intelligent diagnosis method of surface discharge based on visible image and machine learning is proposed in this paper. Firstly

a visible image library of surface discharge under different experimental conditions was established in a dark environment

and the chromatic

gray and morphological features of images were extracted. The image library was divided into 4 stages: normal

slight

serious and dangerous by an unsupervised algorithm. Combined with spectrum correlation experiments

the physical significance of the division results was explained. Four classic supervised learning algorithms

namely

k-nearest neighbor

decision tree

support vector machine

and gradient boosting decision tree

were used to train intelligent diagnostic models by staged datasets. Besides

artificial neural networks were used to compress and reduce the dimension of original data and extract talent features for recognition. The results show that the supervised learning model based on color features has a higher accuracy for discharge diagnosis; the accuracy of artificial neural network reaches 0.958

which is obviously higher than that of classical learning algorithms（the accuracy was 0.886）. The above results show that the visible light image can be used to diagnose surface discharges effectively. The application of machine learning ensures the efficiency and accuracy of diagnoses. This method will enable status identification and fault location to be completed simultaneously

greatly reducing costs

and providing a strong guarantee for the safe and stable operation of the power system.