THERE are 4000km of railway tunnels in Japan, and with more than 70% now over 60 years old life-extension measures have become a necessity. At present, there are sufficient numbers of qualified inspectors able to assess the condition of tunnel linings, but with the active workforce expected to shrink over the coming years due to an aging population, the need to streamline inspection processes has become a pressing challenge.
Under the current performance-based maintenance system for tunnels, defects are first identified visually. Detailed investigation is then conducted of defects that raise significant concern, and structural soundness is evaluated based on cause analysis and performance verification. If performance requirements are not met, or if there is a recognised possibility that they will not be met in the future, appropriate measures must be taken. In addition, the general inspection, conducted in principle every two years, plays an important role in understanding structural condition and capturing environmental changes.
With support from the Ministry of Land, Infrastructure, Transport and Tourism under the Transportation Technology Development Promotion Programme, RTRI has developed two technologies that make use of continuous tunnel lining imaging technology, which has been widely adopted in recent years as an alternative to manual sketching.
Defect detection
The first is a defect detection and soundness assessment app, which uses an AI agent to automatically detect defects and repair marks on images of tunnel linings, assessing the structural soundness of the tunnel and identifying critical areas requiring on-site verification. To streamline on-site verification, RTRI has also developed a mobile critical area projection system which projects the location of defects directly onto the surface of the tunnel lining.
In general, AI detection technologies demonstrate high accuracy when the target is easily identifiable, as in the case of facial recognition. However, as tunnel defects often overlap and are difficult to distinguish, it was expected that only moderate accuracy would be achieved. The current process for identifying and logging the location of defects relies entirely on manual labour, requiring considerable effort. Considering that even moderate accuracy would be useful as a support tool, this was set as the objective of the first stage of automation and the system was developed on the basis that inspectors would review and amend the results.
Targeted defects included exposed rebar, rust stains, and water leakage, which provide critical information for soundness assessment. Deposits such as efflorescence and free lime were also included as they can serve as clues for detecting cracks. To provide training data for the AI system, a bank of images collected by RTRI over a period of several years was used. As the accurate annotation of defect locations on images requires expert knowledge to distinguish between different types of defect, all annotation was performed by railway tunnel specialists. Consequently, partial lining repairs and drainage pipe installations were also targeted for detection as repair marks. The model achieved a success rate of 90% for exposed rebar and over 95% for the other defects.
For soundness assessment and identifying critical areas to be verified on site, an assessment matrix was created. Under the current maintenance standards for railway structures in Japan, during a general inspection of a tunnel structural soundness is typically categorised as A, B, C or S, the most severe condition being category A. In developing the defect detection app, these four categories were used to assess tunnel structural stability, chosen as the focus among performance criteria. Critical areas and soundness are assessed to produce a conservative, safety-focused evaluation which reflects published RTRI guidelines on tunnel assessment, including locations where hammer tests should be conducted on site.
As can be seen in Figure 1, the critical area projection system is installed in a lightweight trolley which is pushed by hand along the track. Images are automatically and continuously projected onto the correct position in the direction of travel, while a rotating base allows the projector angle to be adjusted, enabling the mesh diagrams to be adjusted and projected along the circumferential direction of the tunnel as well.
Critical area projection software is installed on a tablet computer. The software creates mesh diagrams that highlight the automatically identified critical areas, adjusting them according to the position along the track relative to kilometre posts, tunnel lining curvature, and projection distance. The mesh diagrams are position-corrected using a device that measures the distance travelled.
The shape of the tunnel and the clearance from the track centre to the lining surface vary from site to site, requiring projection corrections according to local conditions. The projection system is designed to perform these corrections using software, enabling the use of commercially available projectors.
Figure 2 shows projection in progress in an actual tunnel, with square meshes correctly projected according to the cross-sectional shape at that particular location. During testing, the accuracy of the system used to measure the distance travelled by the trolley was verified over a 400m section within the tunnel. A high-precision rail distance meter was used to measure the actual distance travelled. The results indicated an error rate of 0.18% per 100m.
The alignment of the defect locations with the red portions indicating critical areas also confirms that a level of accuracy in distance travelled is maintained. Tunnel inspection work during overnight possessions is limited to around 3-5 hours, with the result that only a few hundred metres of tunnel lining can be inspected each night. As the number of adjustments required each night would be minimal, the accuracy in measuring distance travelled was deemed sufficient for practical use.
With the objective of making further improvements to reduce the time and labour required to inspect tunnels, RTRI is now building on its work on automatic soundness assessment and critical area projection to develop a system that is also capable of projecting the results of hammer test inspections on the tunnel lining. This involves recording the sound of a hammer test and using AI to distinguish between solid and hollow sounds. Simultaneously, the position of the hammer is tracked using a CoreHW locator and tag to display the results on the tunnel wall in real time. The objective is to reduce the variability of manual inspection and eliminate the risk of overlooking locations where tests should be carried out and attention is required.
*Takashi Nakayama is head of the Tunnel Engineering Laboratory within the Structures Technology Division of Japan’s Railway Technical Research Institute (RTRI). Keisuke Shimamoto and Kazuhide Yashiro are both senior researchers at the laboratory.
