Automated Corrosion Detection on Steel Structures Using Convolutional Neural Network

Abstract

Steel is a material that is widely used in industry and construction. The tensile and compressive force of steel is relatively high compared to other materials. On the opposite, low corrosion resistance is the main weakness of steel, which can encourage steel deterioration and fatal accidents for the user. Furthermore, regular visual inspection by a human should be performed to prevent catastrophic incidents. However, human visual inspection increases the risk of work accidents and reduces work effectiveness. Therefore, a drone with a camera is one solution to increase efficiency, increase security levels, and minimize difficulties or risks during corrosion inspection. In this research, the drone has been used to capture corroded video of a construction structure. The convolutional neural network (CNN) method is then used to detect the location of the corroded images. This study has been conducted on Surabaya’s Petekan-bridge with the Mobilenet V1 SSD pre-training model. In this study, the distance between a drone and the detected object varied between 1 and 2 m. Next, the drone speed was varied into 0.6 m/s, 0.9m/s, and 1.3m/s. As a result, CNN can detect corrosion on the surface of steel materials with the best accuracy is 84.66% and minimum total loss value of 1.673 by applying 200 images, 200000 epochs, batch size at 4, learning rate at 0.001 and 0.1, the distance at 1 m, drone speed at 0.6 m/s.