Automatic detection of cracks using deep learning has been studied as one of the measures to cope with the aging of concrete structures, which is currently a problem throughout Japan. In addition, research is also being conducted on width estimation to measure the degree of damage of cracks to assist inspection work. In this study, we used probability maps of cracks obtained from deep learning to estimate widths with higher accuracy and to calculate confidence levels corresponding to the estimated widths. The statistical analysisof the error between the estimated width and the correct width in 17 images for evaluation showed the relationship between the confidence level and the error of the estimated width and confirmed the validity of the results.

In order to promote digital transformation (DX) at civil engineering construction sites, efforts are being made to diversify and combine new technologies. The objective is to optimize the construction pro-duction process by reforming the business process, to solve the problem of human resource shortage and asset management of infrastructure assets, and to realize sustainable productivity improvement. How-ever, introducing those new technologies alone is not enough,new issues are being uncovered, such as the occurrence of human error and sluggish migration from legacy systems. The key to countering those problems are to integrate the skills of skilled workers with new technologies.

In particular, the author believes that the decision-making of skilled workers should be modeled through the use of data. In this paper, we introduce the implementation of the model by building a system based on trial data at the actual construction site and propose the way of operation for the next generation from the viewpoint of the construction site.

The infrastructure industry faces a variety of pressing issues, such as the aging of the population and the shortage of human resources due to a shrinking workforce, that need to be solved by DX technology. In particular, laser scanner and point cloud analysis technologies have evolved remarkably, and it is now possible to convert the surrounding environment into a point cloud with high accuracy and in a short time, and automatically calculate the desired values. In this paper, we focus on the flatness inspection of concrete slabs of bridges, and develop a technology to instantly calculate the height of slabs by using a ground-based laser scanner and automatic point cloud analysis technology with Infracapture, instead of the conventional method of measuring the height of slabs at a huge number of measurement points.

The developed system was compared with conventional methods and evaluated in terms of accuracy and inspection time. As a result, the system succeeded in reducing measurement time by 88% while ensuring accuracy, and was found to make a significant contribution to the reduction of physical workload.

In this study, we propose a method to predict locations of road-related events from urgent call data by using large language models. Operators need to identify the location of road-related events from information verbally conveyed by the reporter during the call, and this requires them to have both job experience and geographical knowledge. We aim to construct the framework that predicts the location from urgent calls to alleviate the burden on operators in their operations. Thus, we utilize the large language models with extensive pre-training knowledge to extract location information from the text transcribed using a speech recognition model. We evaluate the proposed method using real urgent call data to assess its effectiveness and highlight the remaining problems of this study.

In order to improve the quality of maintenance management against scouring disasters, which threaten the safety and stability of railroad transportation, it is necessary to evaluate the risk of localized scouring at the piers of railroad river bridges and to identify river bridges with a high possibility of being damaged. To this end, a database was constructed by collecting and organizing conditions related to river characteristics and bridge structures from past cases of excavation damage and unaffected cases. We also attempted to construct a learning model that determines whether a bridge is damaged or not by using a decision tree algorithm, which is one of machine learning models, on the database. The prediction results obtained from the learning model were compared with the evaluation scores from the scour scoring table, and the usefulness of the machine learning model was verified.

Although regional railway companies are facing severe business conditions, especially due to the declining birthrate, aging population, and the recent coronavirus pandemic, they need to properly inspect and maintain railway facilities and rolling stock to ensure safe and stable operation of train. In this study, we developed a train patrol support application based on the requirements specified in the "Maintenance and Management Standard for Railway Structures (Track part)", aiming at the practical application of a train patrol support method using a smartphone as a low-cost track condition management method that can be introduced even by regional railway companies. Next, we conducted test measurements using the developed application on the commercial lines of several railway operators, and investigated the utilization of the measurement data, such as acceleration and forward view video, obtained from various angles. As a result, the acceleration data is expected to be effective for train vibration management, and the forward view video data is expected to be effective for understanding track conditions when reviewed at a desk.

Natural language processing (NLP) is expected to be one of the interfaces to strengthen the tight connection between physical and virtual spaces in the digital twin, and is being considered for practical use in the field of civil engineering. For this technology to be fully functional, it is necessary for the language models to understand civil engineering terminology in the context, and appropriate evaluation of the models is required. However, most of the previous studies have focused on how to adapt the technology to the civil engineering field, and have not focused on the evaluation of the capability of language models to generate sentences. Therefore, this study aims to establish an evaluation metric to evaluate the capability of language models in the field of civil engineering. As a first step, we created a new dataset for evaluation and compared which of the existing metrics are appropriate for the civil engineering field. Finally, we discuss the issues involved in considering an automatic evaluation index.

In recent years, efforts have been made in Japan to create walkable street spaces. In this effort, it is important to investigate and understand the action of users of street spaces. However, performing these tasks manually is burdensome in terms of both time and cost. In this research, we are developing a human action detection technology using deep learning, with the aim of improving work efficiency. In this paper, we developed a model that detects various actions listed in the guidelines prescribed by MLIT using camera images taken of street spaces. Then, we showed that it is possible to quantitatively understand the usage situation of actual street space.

In 2023, the Third National Land Formation Plan (National Plan) was decided by the Cabinet; revisions of local plans in various regions based on this plan are anticipated. This study examined the potential for recent artificial intelligence (AI) utilization in national land development, with a focus on human-centered design that considered the opinions and requests of the residents. The results of a quantification theory analysis of previous national surveys regarding social infrastructure suggest that residents’ perceptions of current social infrastructure development vary according to living area and age. An experimental examination of urban landscape images of major and local cities using recent AI revealed that the use of large-scale language models and layer-generation AI may be effective in developing urban plans that are innovative, dynamic, and most importantly, human-centered.

Skill transfer to young engineers from senior engineers is a very important task in infrastructure equipment inspection. To support the skill transfer, an analysis method of the key factors of senior engineers skill is needed. However, conventional research has been limited to skill-level classification or analysis of the relationship between the skill level and biological data such as eye gaze and motion obtained from the engineers. This paper presents a method of classifying the skill level and visualization of its key factors to support the skill transfer. The proposed method employs a graph convolutional network introducing a novel attention mechanism for the classification and visualization.

In this study, we propose a zero-shot high-risk situation detection method based on object detection and pose estimation using a fixed camera, aimed at preventing labor accidents at construction sites. By utilizing pre-trained object detection models and pose estimation models, our proposed method determines two high-risk situations: when a worker is unaware of an approaching vehicle, and when a worker is in the roadway. Our proposed method enables the detection of high-risk situations and is expected to enhance safety awareness among workers and managers by providing statistical information such as the time of occurrence and frequency of the detected high-risk situations. At the end of this paper, we verify the effectiveness of the proposed method through experiments using the videos taken at the entrance and exit of a construction site.

In this paper, we propose a method for estimating the contact accidents risk with heavy machinery to support safety management on construction sites. According to recent reports on occupational accidents, since the construction industry experiences a high number of incidents, preventing contact accidents between heavy machinery and workers, which are on the increase, is a crucial task. The proposed method constructs a deep learning model to estimate the contact accident risk by using videos obtained from multiple viewpoints of cameras mounted on heavy machinery or fixed-point cameras at a construction site. By inputting visual information of videos obtained via spatial-temporal attention to a recurrent neural network, it is possible to accurately estimate the risk of contact accidents. At the end of this paper, we can verify the effectiveness of the proposed method through experiments using videos taken at actual construction sites.

Digital twin technology, which reproduces real phenomena in virtual spaces, is becoming increasingly important in various fields of civil engineering. Especially for the future prediction of rivers, which are long natural structures, the availability of information that canbeprocessedautomatically iscrucial for sustainable infrastructure maintenanceamidstadeclineinengineers. This paper demonstrates a module developed towards realizing digital twins, focusing on riverbed variation analysis used for medium to long-term monitoring and trend prediction within river channels. As a practical application example, the module was demonstrated using various river data of the Kamanashi River, including 3D terrain data, to automate acquisition, conversion, analysis, and distribution. The module is designed to be customizable and replaceable in response to technological advances, making it a characteristic contribution to the automation of data processing in the future.

Improvements in measurement technology have enabled the acquisition of large-scale data, leading to the growing interest in the concept of Digital Twin. In Digital Twin, data science plays a crucial role, and Dynamic Mode Decomposition with Control (DMDc), a data-driven approach, has gained attention in recent years. The authors have previously explored the applicability of DMDc in predicting future settlement in a reclamation area. In this study, the method is improved by applying DMDc with a formulation of the Robbins-Monro algorithm and Kalman Filter to predict settlement. This paper demonstrates the potential improvement in prediction accuracy in the prediction period when observational data is available for some monitoring points after the learning phase. Additionally, it provides an example of quantifying the uncertainty of predicted settlements, which was not achievable with DMDc alone.

Periodical visual inspection has been started in Japan. According as the amount of inspection data increases, consistent and traceable data management is crucial. Digital transformation plays a very important role. We are developing 3D models in which inspection data can be saved. Additionally, we are also developing structural 3D model by the finite element analysis, integrating the monitoring data for the smart infrastructure management. An existing bridge was picked up in Tokyo area. Comparison between finite element analysis and field loading was made. Agreement between field loading and analysis is very important because of the simulation of deterioration on the structure is very crucial. The influence of appendages such as fense, covering plate to the stiffness of the structure was not negligible. Considering those, the accuracy was improved up to 92.7%. The error of the analysis was corresponding to 1 MPa by stress value. By the use of numerical model of finite element analysis, prediction of the weak points is the next step.

The Ministry of Justice has been promoting the Cadastral Map Creation Project since fiscal year 2015 to clarify land locations and boundaries. In January, 2023, the nationwide cadastral map data was freely available to the public through the G-Space Information Center. However, owing to the fact that much of the various data is based on arbitrary coordinate systems, this has become a challenge for data development and utilization for various purposes. Therefore, this research proposes a method to automatically convert cadastral map data from arbitrary coordinate systems to a public coordinate system by referencing linear and attribute information and using geocoding and alignment techniques. As one use case for the trans-formed cadastral map data, a technology to estimate block wall ownership is proposed, with the extent of its utility confirmed through practical experiments.

With the use of laser measurement equipment, point clouds of national land space have been measured and accumulated. Since point clouds, the product of laser measurement, is a collection of points with positional coordinates, it must be processed into a data set that can be easily processed spatially to be deployed for a variety of purposes. The authors have been working on the product model of point clouds and have constructed a road feature identification model using area data generated from road drawings and road maps for automatic driving (HD maps) and point clouds for each road feature. On the other hand, the problem was that the area data necessary for product model could not be generated for sections where there were no drawings or maps.

In this study, we devised a method to automatically generate the area data of landmarks using the land- mark identification model. We compared the area data of HD maps and the proposed method and found that the proposed method can automatically generate area data from point clouds of road feature.

Toward the expansion of the digital twin study of wide-area expansion and service cooperation of city OS Yoshihiro NAKAGAWA and Toshihiro KAMEDA We examined wide-area and heterogeneous service cooperation for more efficient data linkage and utilization using a city OS toward the realization of a smart city for society 5.0. We used FIWARE, which is widely used as the city OS. LoRaWAN, one of the LPWA (Low Power Wide Area), was used for wide area coverage. The Things Network, a network of LoRaWAN, was used for examples of service cooperation. This was compared and examined with the case of the MILT Data Platform as another example of other service cooperation. It is necessary to design a flexible conversion method of data structure, commonization, and database arrangement depending on the frequency of access to services.

To build a digital twin, we devised a microservice that aggregates urban environmental data to TTN service using LoRaWAN, and further distributes it using GraphQL and REST API. The use of LoRaWAN and TTN makes it easy to build a sensor network. In addition to the currently popular REST API, we use GraphQL, which has been adopted by MLIT Multi Data Platform, as the delivery method. GraphQL has the advantage of allowing users to specify and flexibly format the data to be retrieved, and is expected to expand the range of client implementations. The system was verified through actual construction, showing that it is feasible and clarifying problems.

In the maintenance work of steel bridges, repairs and reinforcements are carried out to improve performance deterioration due to corrosion or fatigue, or to meet the requirements of the latest standards. Repairs and reinforcements often involve adding new members such as patch plates to existing members, or replacing members. It is necessary to understand the detailed shape and dimensions by measuring the existing members, as well as to conduct a three-dimensional examination of the structure and construction method based on sufficient knowledge and experience. Aiming to improve productivity, we focused on easy-to-understand three-dimensional augmented reality technology, which allowed us to image new members by directly seeing the existing members on site. We considered incorporating into practice a method to confirm the installation status in three dimensions in advance. As a result, we clarified that it can be implemented easily with sufficient accuracy.

In winter road environments of snowy and cold regions, drivers face severe conditions such as icy sur faces, low visibility, and roads narrowed by snow accumulation. Current strategies to aid road users mainly provide information about road conditions, snow coverage, and visibility. However, this paper introduces a novel approach focused on driver behavior, particularly changes in the lateral position of vehicles, rather than road-focused information. High-stress and dangerous driving environments often force drivers to alter their driving behavior. By visualizing locations on a map where lateral position changes frequently, we aim to assist drivers in selecting safer routes. This paper proposes a method to visualize these changes in driving position using GNSS data and demonstrates the effectiveness of this method through experiments.

General object recognition can be used to recognize traffic participants by applying it to images of road spaces. However, general object recognition, which identifies objects based on their appearance, has a problem with the detection and recognition of two-wheeled vehicles. In this paper, we propose a method to identify two-wheeled vehicles by applying general object recognition and human pose estimation to road videos, focusing not only on the vehicles themselves but also on the drivers of the twowheeled vehicles. The proposed method first applies general object recognition and human pose estimation to road video. Next, we track the obtained posture data and obtain features that represent the motion of the two-wheeled vehicle driver. By constructing a discriminator using the obtained features as input, a two-wheeled vehicle driver is extracted from the results of human pose estimation. Next, the extracted two-wheeled vehicle drivers are discriminated into bicycles and motorcycles using another discriminator. By integrating the identification results based on human pose estimation with the results of general object recognition, it is possible to identify two-wheeled vehicles that could not be recognized by general object recognition and thus to improve the accuracy of the system. In the last part of this paper, we confirm the effectiveness of the proposed method through experiments using actual road videos.

The data obtained from various sensors is utilized for spatial analysis utilizing AI. To maximize the effectiveness of AI-based spatial analysis, it is necessary to estimate in advance factors such as the placement, installation angles, and quantity of multiple sensors. To make this possible, a search method for camera placement conditions based on iterative deepening depth-first search utilizing an evaluation function based on object detection results was constructed. However, there is a challenge of falling into local optima due to the limited search space. Therefore, in this paper, a search method for camera placement conditions based on genetic algorithms is proposed. Specifically, the solution obtained by the conventional method is used as the initial solution, and selection, crossover, and mutation are iteratively applied to explore camera placement conditions. This allows for searching broader search compared to conventional methods and achieves a method less prone to local optimal solution. The effectiveness of the proposed method is confirmed through experiments.

In Digital Twin, physical space and virtual space are connected by data obtained from sensors and other sources including by human operators. It is not realistic to detect data with different results before and after processing due to human intervention by multiple human checks. This study verified whether ChatGPT, a chatbot using a large-scale language model, can detect such discrepancies in data. This experiment used local government data from xROAD’s road bridge data. The results show that ChatGPT may be able to detect discrepancies between the data obtained from the public API and the inspection report with No Code.

A drone equipped with a scanner that emits green laser that penetrates underwater has been put into practical use, and it is expected that it will be used to improve the efficiency and sophistication of river management using 3D data. This study focused on the application of this river management to survey of damaged areas or disaster recovery, and examined effective ways to use drone surveying. Specifically, we will verify the accuracy of surveying when carried out without adjustment points, and also establish a process to quickly detect locations where deformations occur from survey data obtained before and after the flood, and quantify their magnitude due to developed algorithm. This paper describes on the application of drone surveying that meets the required accuracy for use in disaster response along with examples.

The efficiency of reinforcement confirmation in reinforced concrete structures was aimed to be enhanced by verifying the effectiveness of a camera attached to a rebar tying machine that automatically moves along the slab-shaped reinforcement and captures videos. The obtained rebar model from the point cloud generated through Structure-from-Motion (SfM) and Multi-View Stereo (MVS) was utilized for confirmation tasks such as checking the spacing of reinforcement. Since the machine can move regularly on the slab, capturing videos from multiple perspectives, it efficiently acquired point clouds of the lower-level rein-forcement behind the upper-level rebar and the point cloud of the stirrups arranged vertically. Additionally, using the generated cylindrical rebar model from the obtained point cloud, it was possible to measure the spacing between all the reinforcements within the target range of the structure.

In recent years, the increase in extreme weather events has caused serious social problems such as increased risk of heat stroke. As a countermeasure, the Earth Simulator supercomputer is being used to simulate heat, which reproduces temperature, humidity, and other environmental factors. Since the prediction of the heat environment on an urban scale is required, it is necessary to ensure high accuracy through data assimilation. In this study, a sensor network system that assimilates measured data from physical space was constructed to improve the accuracy of the simulation and to realize a digital twin of the city through data assimilation. As a result of a demonstration experiment conducted in Yumeshima, the site of the Osaka Expo, it was confirmed that the sensors and communication worked properly and contributed to the construction of a digital twin.

This study reports on a new system for predicting a risk of tipping aimed at improving the safety of remotely operated hydraulic excavators. In disaster recovery sites, remotely operated hydraulic excavators are sometimes deployed, but operating them involves risks, including the danger of tipping over. To prevent the risk of tipping, a method of predicting the danger of tipping in advance is effective. Existing prediction methods based on physical models do not consider external disturbances or vehicle movement and the time taken for prediction has been a challenge. In this study, we propose a new system that uses a DNN (Deep Neural Network) to quickly predict the maximum inclination angle of a hydraulic excavator up to one second ahead. This system does not require the analysis of complex physical models and directly learns and predicts the relationship between sensor data and future body inclination angles. Furthermore, by converting 3D point clouds into bird’s-eye views and inputting them into a CNN (Convolutional Neural Network), we aim for rapid prediction. Verification by simulation showed a prediction error of 0.056 radians and a prediction time of about 2.79 milliseconds, demonstrating sufficient performance. However, these results are based on simulations, and validation with actual excavators is a future challenge. This study is expected to improve the safety of hydraulic excavators operated remotely. In the future, the goal is to achieve more efficient operations while ensuring the safety of both operators and equipment.

The purpose of this study was to confirm the accuracy of 3D topographic point cloud data created by combining smartphone LiDAR and ichimil, a network RTK positioning service, for the purpose of inexpensive and simple creation of 3D topographic models for small-scale earthwork projects. The coordinates of the simplified points were separately measured by ichimil, and these coordinates were added to the 3D point cloud data acquired by smartphone LiDAR. The 3D point cloud data was created by UAV photogrammetry for comparison, and the accuracy of the 3D point cloud data created using both smartphone LiDAR and ichimil was checked against this data. As a result, it was confirmed that the accuracy was within 200mm in both the plane and vertical directions, satisfying the accuracy requirements of the partial payment form survey.

With progress of Web production technology, websites are being mass-produced every day all over the world. In the context of roadside stations gaining attention as hubs for regional revitalization, there is a current reevaluation of the creation and operation of websites for each roadside station, addressing the challenge of insufficiently showcasing the diverse charms of different regions. As users commonly browse websites with higher search rankings in web searches, it becomes crucial to ensure that the web-sites of roadside stations rank prominently in search results. Therefore, it is difficult to formulate coun-termeasure policy for own website. In this research, we propose a present method to evaluate SEO measures items from relation to search ranking by analyzing many applicated web pages comprehen-sively. These web pages retrieved from web search engine using sets of search queries related to main keywords of website.

In the development of Business Continuity Plans (BCPs) for social welfare facilities, establishing clear criteria for emergency staff gathering is crucial in ensuring the safety of both users and staff members. However, it is difficult to utilize traditional hazard maps for imagining specific disaster scenarios on a timeseries basis and formulating corresponding plans. In this study, crowd simulations were conducted for BCP formulation, assuming flooding due to river overflow. These simulations enabled us to understand the times and routes that allow staff residing in each area to safely access the facility. This suggests the possibility of a new approach to BCP formulation, especially for more concrete planning of the optimal timing and methods for the emergency staff gathering. In the future, we will conduct workshops on BCP formulation at social welfare facilities using the results of this study and evaluate their usefulness.

In general, the scale of an earthquake disaster is often expressed in terms of the maximum seismic intensity. However, in considering the impact on the lives of disaster victims, it is necessary to take into account the distribution and composition of the population. To address this issue, we developed a series of methods to calculate the population by seismic intensity by combining the estimated seismic intensity distribution published by the Japan Meteorological Agency and census mesh statistics, and actually applied them to the 2024 Noto Peninsula earthquake. As a result, we were able to quantitatively analyze the effects of a large-scale earthquake that occurred over a wide area, and to obtain a quantitative and comprehensive understanding of the extent of the earthquake’s impact and its bias. The use of mesh statistics made it possible to easily analyze various types of data (e.g., by attribute) with a small amount of computation, making it a method that can be used quickly immediately after a disaster.