Korean researchers unveil AI-powered inspection drones that generate lifelike damage data to tackle tunnel safety — revolutionizing how aging infrastructure is maintained without the need for massive datasets.
This is a training structure for a crack segmentation model using a generative AI approach. This method enables the detection algorithm to adapt to new environments.
Korea Institute of Civil Engineering and Building Technology has developed 'Generative AI-Based Inspection Technology' to safely construct and maintain urban underground highways.
Recently, the number of aging tunnels has been gradually increasing. However, the number of specialized personnel who can manage and inspect them is decreasing. To address this, practical measures incorporating IT technology are urgently needed. Developing a high-performance AI model requires a vast amount of training data. However, when applying such AI technology to maintenance sites, there is a challenge of data scarcity and field adaptability. Deep learning models require a large amount of training data, but it is difficult to obtain data as damage scenes such as delamination or rebar exposure on concrete surfaces are not commonly seen. To solve this issue, small data learning using a small number of field images rather than big data learning with concrete damage images is necessary.
In response, KICT research team led by Dr. Shim, Seungbo, at the Department of Geotechnical Engineering Research has developed a 'smart' AI inspection technology that overcomes the existing limitations for the safety inspection of aging tunnels. The most notable feature of this technology is its ability to synthesize unique concrete damage scenes seen only in aging infrastructure, even with a small amount of data. Previously, collected field data were processed to detect cracks, but the newly developed generative AI can generate data so sophisticated that it is indistinguishable from actual footage. The AI has the capability to synthesize 10,000 images of concrete damage within 24 hours, and through adaptive technology that learns and trains detection models based on collected field video data, it has effectively addressed data scarcity issues and reduced training costs.
This AI technology has been integrated with autonomous drones, successfully completing field verification within actual large-scale tunnels. The most critical part of tunnel inspection is the ceiling. Currently, workers perform visual inspections using high-altitude work vehicles, but this method may have reliability and safety issues. However, the drone developed in collaboration with LASTMILE Co., Ltd. (a KICT Resident company) can freely navigate inside tunnels with a margin of error of 20cm, using a 200M-class long-range indoor positioning sensor. This is expected to effectively replace the workforce performing hazardous tasks.
Dr. Shim said, "This research is a technology that breaks the stereotype that a large amount of training data is required to utilize artificial intelligence, and implements a new concept of creating data if it is not available." He expressed his expectation that this technology will open new possibilities for AI applications across the entire construction field.
The research was conducted backed by National Research Foundation of Korea (NRF) [2022R1F1A1074663] and was carried out under the KICT Research Program (project no. 20240051-009, Development of High-Performance UWB-Based Small AI Drone Navigation Technology for Tunnel Safety Inspection) funded by the Ministry of Science and ICT.
AI Reshapes Customer Service: Big Gains for Some, Trade-Offs for Others