Study on the Applicability of AI to the Face Observation of Mountain Tunnel

Introduction

  In the construction of mountain tunnels, support patterns and auxiliary methods are determined after comprehensively observing the state of weathering, cracking, and other conditions of the bedrock that appear at the farthest extent of excavation (called "observation of face ") and then grading those conditions (Fig. 1). However, observation of face have any various challenges, including differences in judgment that arise depending on the individual engineer's experience and cases when making a judgment is difficult. On the other hand, leading to instances and studies involving the use of AI (e.g., image analysis technology, etc.) in observation of face are increasing by recent advancements in AI technologies. Yet the reliability of AI-based face observation and conditions for its application have not been established and many uncertainties are thought to exist here.In light of this, we are conducting a study on possibilities for face image analysis using " deep learning " to clarify AI's applicability to face observation.

  Fig. 1: Engineers conducting face observation using a face observation chart

Experimental Application of AI in Face Observation

  To analyze face images using deep learning and determining the state of the face, we have the deep learning system learn face images and face observation records as training data. Here, deep learning employs a convolutional neural network (CNN) that is used in image recognition (Fig. 2).

  As a test, we prepared a learning model using hundreds of face images that were labeled with the scores provided under face observation item " E:Crack interval." Specifically, we divided the face into three parts as shown in Fig. 3 and then evaluated the face with respect to each region. We created AI that used this model to evaluate crack intervals and we Validate prediction accuracy by inputting new face images. The validation results matched with precision of about 60 to 70%. However, the following problems became apparent from the test results.

(1) AI's decision-making process is unclear.

(2) The specifications of face image data are not standardized at each site and therefore pre-processing is extremely labor intensive.

(3) Scores are biased in training data.

  Particularly with regard to (1), the decision-making process is not revealed, which makes it difficult to use it as foundational material in, for example, discussions concerning a change in agreement between ordering and supplying parties. Moreover, in actual observations of face conducted by engineers, the engineer makes a comprehensive decision using various forms of information, such the conditions of muck and spring water and continuous change. Accordingly, it is thought that there are certain limits in making judgments about a face using just face images at the present time.

  For (2) and (3), a vast amount of data must be collected simply and in a standardized manner at different sites having varying conditions. Moreover, training data contains decisions made from an engineering standpoint and therefore cannot be assumed to be 100% accurate. These and other factors demand thorough study.

Fig. 2: Conceptual image of CNN (from Haruka Adachi [2018])	Fig. 3: Face image used in learning

Conclusion

  The Tunnel Team is currently studying the conditions such as face image data that is effective for study of AI. We are also studying the usefulness of various types of measurement data as training data. We are currently at the test stage and believe that sufficient study will be required before AI can be applied to actual worksites. We intend to make further studies on possibilities for applying AI to face observation by, for example, acquiring experimental face images at actual worksites and studying AI of varying configurations.

(Contact: Tunnel Research Team)

Salmon Spawning Habitat Rehabilitation on Toyohira River, Sapporo, Hokkaido

  The release of artificially hatched chum salmon (Oncorhynchus keta) fry has been conventionally practiced in Toyohira River, Sapporo, Hokkaido. However, half of the chum population returning consists of wild fish. In order to preserve wild chum salmon populations, it is important to improve spawning habitat environment.

  While salmon spawning has been observed in Toyohira River, the spawning habitat environment has degraded. In an alcove downstream of the sandbar located in the middle of the river, the accumulation of fine sediment caused a decrease in the number of redds (until 2016) (Figure 2). In order to address this issue, in collaboration with the Hokkaido Regional Development Bureau and other organizations, we constructed a channel from upstream to the alcove and investigated the number of salmon spawning beds(Figure 3). As a result, the sediment thickness reduced to less than 5 cm (Figure 1) and the number of spawning beds increased after the establishment of the channel.

  A further detailed analysis revealed that the number of spawning beds of the earlier populations, which migrated the river from September to November, and the populations that is late in migration thereafter, both increased (Figure 2).

  The research area was located at the toe of the Toyohira alluvial fan, where groundwater often occur and bring warmer water than the surface water. These warm ground water and the decrease in fine sediment accumulation contributed to the increase of spawning beds of the latter populations in the alcove with the channel.

  Meanwhile, riffles and pools were formed in the channel, where the pebble-cobble stream bed with high water flow velocity were observed. In this area, subsurface water temperature similar to the surface water temperature seems to have been formed, creating an ideal spawning environment for the earlier salmon populations.

  The increase in the number of spawning beds of both groups of populations was a result of the rehabilitation of such spawning environment, which was achieved by the construction of the channel.

  We thank our collaborators in this study, Dohkoukensetsu, Co., Ltd., Sapporo Salmon Museum, Sapporo Wild Salmon Project, and the Sapporo River Office, Sapporo Development and Construction Department, Hokkaido Regional Development Bureau.

Figure 3. Alcove and the constructed Channel in the research area(Captured in June 2018)

(Contact: Watershed Environmental Engineering Research Team, CERI)

#Go to Top of This Page

A Study on Prevention of Re-degradation of Surface Protection with Continuous Fiber Sheets

Photo 1. Anti-seismic reinforcement of bridge
piers using continuous fiber sheets

  One of the anti-seismic reinforcement methods for bridge piers is to wrap them with continuous fiber sheets. A continuous fiber sheet is a sheet of material woven from fine carbon and aramid fiber strings. Unlike other anti-seismic reinforcement measures, this method can be carried out manually without using heavy machinery, as it uses lighter materials. Another advantage of using this method is that the completed work seldom interferes with the flow of the river thanks to its thinness. However, some materials used for continuous fiber sheets are vulnerable to ultraviolet rays. Therefore, the applied sheets are typically covered with protection mortar to save their surface from outer degrading factors.

(1) Falling of exfoliating parts	(2) Collision with flood debris	(3) Abrasion by ocean waves

Photo 2. Examples of Exposure of Continuous Fiber Sheets

Photo 3. Example of an Exfoliated
Part Extending along a Crack

  In recent years, exposure of continuous fiber sheets has been reported at some of the bridge piers reinforced with this method, as a result of the aging, peeling, and loss of surface protection mortar used to protect the sheets from outer degrading factors. Exposure poses a risk of degradation of the sheets by ultraviolet rays and rupture of the sheets from collision with floating debris, and may result in the loss of the sheets' anti-seismic effects. Therefore, it is an urgent issue to establish methods for improving the durability of surface protection works and for detecting the signs of exposure during inspections.

  In Hokkaido Prefecture alone, more than ten cases of continuous fiber sheet exposure have been reported. Based on the results of a series of on-site investigations, we found that the causes of exposure can be categorized into three groups, as shown in Photo 2: (1) falling of exfoliating parts, (2) collision with flood debris, and (3) abrasion caused by ocean waves. Among these, (1) falling of exfoliating parts was the most frequent case. In order to study the causes of such case, we observed how the state of exfoliating surface protection mortar changes over time. As a result, it was found that degradation advances at a higher pace in exfoliated parts that contain cracks, such as the one shown in Photo 3.

  For the future, we plan to conduct indoor simulation tests to study the cracking and exfoliation processes of surface protection mortar, so that we can propose materials and installation methods that can effectively prevent peeling and falling. We also plan to study the conditions of exfoliation that tend to result in exposure of fiber sheets, and to propose effective inspection methods.

(Contact: Material Research Team, CERI)

#Go to Top of This Page