Research And Implementation Of The Classification Of Grounding Grid Material Corrosion

Grounding grid is an infrastructure used for grid work, lighting protection and grounding protection, is an important part to maintain the safe operation of the power system. The material of the grounding grid is perennial buried in soil under 0.6-0.8 meters, it is prone to corrosion due to the special environment of soil. In the case of excavation, the previous method of judging the corrosion is to use the human eyes, this method has a great deal of subjectivity. For this purpose, a method using image processing to learn corrosion features automatically is proposed, the method can realize classification and achieve the purpose of judging levels of the grounding grid corrosion material. Thesis work carried out as follows:Firstly, simulate the phenomenon of the grounding grid corrosion in the laboratory, the corrosion images are collected regularly, and a series of pictures with different corrosion degree are obtained. In order to improve the image quality, an algorithm for image enhancement based on the combination of gray scale transform and nonlinear luminance-chrominance color space is proposed, at the same time the traditional Retinex algorithm with color restoration is used to enhance the image as a comparison. Compare the advantages and disadvantages of the two algorithms, and experiment verify the effectiveness of the proposed algorithm.Secondly, a method based on improved artificial bee colony algorithm is proposed to find the best segmentation points, and combines with the seed region growing algorithm to achieve color image segmentation, broaden the training samples, tagged and untagged data sample sets are established to improve the effectiveness of deep learning algorithm.In addition, in order to avoid limitations of manually obtaining features effecting the classification results, we use deep learning model to training samples and achieve a good feature representation automatically. The model is based on the deep belief network(DBN) using bilinear discriminant strategy to obtain a better initial value of DBN, avoiding embedding local minima in DBN training. In addition, the initialization strategy can automatically set the number of nodes in the hidden layer of the DBN network, and improve the effectiveness of the model.Finally, combine two algorithms to construct a classification model. Using SOM network to achieve a rough classification at the beginning, then using K-means algorithm to achieve an accurate classification. Through the test of the classification of corrosion, we can know that the improved model can effectively complete corrosion classification. Then evaluate the protection level and appearance level of the test sample based on GB/T6461-2002.