Realtime damage detection in long conveyor belts using deep learning approach

Japan's Road networks rely heavily on the mountain tunnels due to its topology. During the construction of these tunnels, mucking process is conducted to remove crushed rocks and rubbles in the tunnel with the use of long conveyor belts with the size of 3 to 10km. Regular visual inspections of these belts are carried out tediously by the workers to ensure belt integrity. To reduce the burden on workers, the paper proposes a vision based deep learning solution deployed on an edge device. Proposed framework detects the size of damage ranging from 1 cm to 100cm. Edge device deployment helps the workers to receive the result in real-time regardless of internet availability or working conditions. The effectiveness of the proposed framework is confirmed on 3 tunnel construction sites, with the estimated mean average precision of 85% for crack detection. The study can be applied in other domains of construction industry such as road damage or concrete damage categorization.