界面ラフネスモデルを用いた水道管の腐食画像の生成と生成AIを用いたラフネスの統計的評価

抄録

It is important to perform diagnosis of buried pipes using nondestructive testing and reflect the results in any renewals or repairs. To establish a method to learn data on graphitization corrosion of water main and generate a large amount of data on the thickness distribution of corroded pipes, we create training data under the distribution of the interface roughness model. Using the training data, we build a generative adversarial network, which is one type of generative AI. This allows us to obtain images representing various types of corrosion and generate samples whose statistical properties can be obtained. However, in order to do so, we need to ensure that the samples are accurate and diverse. We propose to generate samples using a conditional generative adversarial network with gradient penalty. This enables the generation of diverse and accurate images and image generation under physical constraints. Here, we compare and evaluate the statistical properties of the parameters of the interface roughness model after resizing and generation by CWGAN-gp for the original image.