Traffic volume surveys are conducted throughout Japan to obtain basic information for road management and other purposes. Traditional manual observation methods have been largely phased out, and various observation methods, such as continuous traffic volume monitoring devices and AI observation using CCTV images, are employed. However, these methods do not provide comprehensive coverage of the extensive road network in Japan. There is a need for a versatile and universally applicable traffic volume survey method that allows for easy installation and removal in any location.

In this study, we developed a traffic volume survey system specifically for sectional surveys. By using easily deployable cameras, edge devices, and LTE router, we were able to acquire real-time traffic volume data on-site using cloud technology. In addition, the system was designed to facilitate data accumulation and utilization. To validate the effectiveness of the system, a traffic volume survey was conducted on the Ube Road as a practical application.

Many organizations have prepared individual support plans to ensure the safe evacuation of people requiring special care, such as the elderly and disabled. This plan enables precise evacuation support to be provided to each individual's needs. However, there are issues such as the aging of the evacuation support personnel themselves and a shortage of evacuation supporters due to depopulation. In this study, we conducted experimental research on the construction of an evacuation support system using autonomous drones for individuals requiring special care who can walk independently. The experiments focused on the evacuation guidance and monitoring of evacuees. As a result, the drone was able to guide evacuees, whether they were alone or in groups, in any direction while maintaining a certain distance from them in the evacuation guidance experiments. In the evacuation monitoring experiments, the drone was able to track the number of evacuees using AR markers and monitor their movements.

In recent years, large-scale language models (LLM) such as OpenAI's ChatGPT have been expected to be used. By using the trained model of LLM, various tasks such as question answering, text generation, translation, summarization, text classification and evaluation can be applied. Although LLM has learned a lot of general texts in advance, As a method of using specialized knowledge, there is a method of extracting features of the target text with LLM and having LLM perform the task again using the features, whereas there is a method of relearning the LLM model itself using original data (fine tuning). In this paper, we will briefly organize the LLM model and organize how to fine-tune the text in the civil engineering field. Using open data such as inspection procedures, we actually performed fine tuning, and from the results, we organized the issues from the viewpoint of training data, utilization of existing models, and evaluation method of generation results.

In this paper, we propose a distress estimation method for road attachments. Road attachments, such as signs and lighting, are equipped over a huge number and wide area, and therefore it is desired to achieve automatic inspection by using drones to reduce the burden on the inspectors. While captured images by drones include the diversity of the background including ground, sky, and road surfaces, the previous methods did not consider the diversity of the background of captured images of road attachments. This study proposes the distress estimation method via attention-based multiple instance learning to address this issue. We input patches of the images into the estimation model to distinguish between the background area and road attachment area and assign importance weight, or attention, to each patch. By performing this strategy, we realize the distress estimation method considering the diversity of the background area of images. In the experiment, we achieve a classification accuracy of about 70 % using images of actual road attachments confirming the effectiveness of this research approach.

In the damage diagnosis of bridge inspections, a comprehensive judgment is made from the inspector's professional viewpoint as well as objective facts such as damage conditions. In this study, we extracted information from inspectors' findings, which are considered to describe the basis for such judgments in damage diagnosis, and used a dataset created based on the extracted information to construct a deep learning model that can answer questions related to the diagnosis of various components and damage in images. The model was able to not only identify the name and type of damage, but also predict the state, cause, and inpact of the damage. Furthermore, the analysis of the model's evaluation and the basis for judgment suggested that the model made judgments based on criteria like those used by humans, such as paying attention not only to the member where damage occurred but also to the surrounding members.

The purpose of this study is to grasp characteristics of each local governments and to utilize for bridge maintenance and management by analyzing of whether existence or distance of bridge detour. Bridge detour is calculated from program that applied the Dijkstra’s algorithm. As a result, it found that tendency of bridge detour condition in each area and correlation with inhabitable area. Some incorrect data found from detour program. However, it can be an indication for understand characteristics of each local governments and social condition of bridge maintenance and management. Online map site constructed with about 23000 bridges detour data displayed on map in Niigata prefecture area.

The utilization of "collaborative AI robot groups" has potential in rapid emergency recovery operations as disasters have recently become more frequent and severe. River channel blockages may cause secondary disasters, which unmanned construction machines need to autonomously avoid. However, that is challenging due to the insufficient availability of AI training data in disaster scenarios. To address this issue, this study provides a workflow for assessing the dangers by integrating AI with human knowledge and experience, which have been accumulated in natural language. First, the analysis results of a landslide site were obtained by using Semantic Segmentation. Secondly, they were transformed into fixed phrases and fed into a large language model (LLM) that had been fine-tuned for landslide disasters to determine the risk. This research exhibits an application example that combines image AI technology, knowledge of civil engineering, and the rapidly developing LLM.

Improving spatial resolutions of precipitation data is one of the most critical issues for disaster mitigations caused by heavy rainfall. In recent years, the deep-learning-based super-resolution attracts interests of researchers for predicting high-resolution image from coarser-resolution input data. This study proposes applying the super-resolution techniques for downscaling spatial patterns of precipitation data. Here we trained two deep learning-based super-resolution neural networks, SRCNN and SwinIR, so that the models predict the original high-resolution precipitation fields from coarsened rain fields. A series of experiments demonstrated that the two deep-leaning-based models outperformed the conventional bilinear interpolation method when sufficient training data are used for the training. Particularly, using SwinIR with the transformer, we successfully achieved super-resolution with higher accuracy than the SRCNN.

In recent years, research such as digitizing the hammering sound and analyzing it using machine learning to identify the damages, etc is active. However, many studies have focused on the determination under ideal conditions in the laboratory, and so the changes in environmental conditions, for example dry and wet conditions of concrete, have not been sufficiently investigated. Therefore, in this study, examined whether it is possible to determine the damage of concrete specimens (RC beams) based on the dry and wet conditions using the local outlier factor method, which is one of the machine learning methods. Specifically, the authors collected sound data from RC beam specimens before and after loading in both dry and wet conditions, and used these data as input for an experimental study. As a result, it was confirmed that there is a possibility that the judgment can be made regardless of the dry or wet condition, depending on the input data.

Data such as pedestrian traffic volume and pedestrian trajectories are used as basic data for urban redevelopment and revitalization planning. In this study, we developed a technology for determining pedestrian trajectories from mul- tiple cameras using Deep Learning and Gradient Boosting, which are artificial intelligence (AI) technologies. Specifi- cally, we improved the accuracy by determining the same person based on the location information of the person de- tected by the Deep Learning model, and then tracking the estimated trajectory using the Gradient Boosting model. As a result of demonstration experiments, the trajectory of a person can be determined even when the person moves across the camera, indicating that this technology may be effective in improving and saving labor for pedestrian traffic volume and pedestrian trajectory measurement.

With the development of AI technology, various inflow prediction methods have been proposed, but to increase reliability and realize social implementation, it is necessary to show the grounds leading up to prediction and ensure the validity of the prediction results. In this study, the target basin is the Gonokawa Haji dam basin, and an inflow prediction model was constructed using a convolutional neural network with the radar-raingauge analyzed precipitation of the Japan Meteorological Agency as the input condition. We applied XAI (explainable AI) technology to this model and visualized and considered the basis of the pre- diction. As a result of the consideration, the validity of the inflow prediction model was confirmed.

In dam management, it is important to provide appropriate water supply based on the water utilization situation such as agricultural water. In the low water management of the Kinugawa River, in order to secure the required flow rate at the water utilization reference point, the flow rate is adjusted by the integrated operation of the four dams and the dam group cooperation facility. Accurate prediction of inflow is necessary. In order to carry out efficient and effective integrated management, this study builds and verifies the accuracy of low water inflow prediction models based on deep learning and tank models that take into account the effects of non-melting snow and snow melting for four dams. carried out. As a result of the accuracy verification, the tank model was partially superior in the non-melting season, but the deep learning model generally showed higher reproducibility. In the snowmelt season, the snowmelt model showed high reproducibility between the non-melt model and the snowmelt model, confirming the effectiveness of the snowmelt model.

The high computational cost can be a bottleneck for ensemble simulations by rainfall-runoff inundation models based on physical processes. To reduce the computational cost, this study has developed a deep neural network (DNN) emulator that can rapidly predict spatiotemporal distributions of inundation depth. The emulator uses the spatiotemporal distribution of precipitation as input data and predicts the spatiotemporal distribution of inundation depth for the same period. Here, the spatiotemporal distribution of inundation depth consists of three dimensions: event, spatial pattern, and time. Therefore, we conduct dimensionality reduction by singular value decomposition as a pre-processing step prior to train the DNN. To achieve this, it is necessary to transform the data from three variables to a matrix format, and we test two different transformation methods. We found that a transformation, which reduced the size of the DNN output more, improved the prediction accuracy. The developed DNN-based emulator showed accurate inundation predictions whose accuracy is equivalent to the one that outputs only the maximum inundation depth.

Currently, the general inspection method for weathering steel is to visually evaluate the rust condition in five levels, however individual differences in judgment are likely to occur. Therefore, a quantitative and automatic evaluation method using a convolutional neural network (CNN) has been proposed. However, there are problems that the rating of a small piece of rust image used as teacher data is unknown. Additionally, the CNN can only judge the rust appearance rating of a small piece of image. In this study, the feasibility of classifying the rust appearance rating was examined by visualizing the features extracted from a CNN that has not been trained on rust images and reducing the dimensionality of the features using an unsupervised learning method t-distributed probabilistic nearest neighbor embedding (t-SNE). The results showed the possibility of classifying the rust appearance rating based on the distribution of the image of small pieces of each appearance rating.

A method of inspecting the condition of utility poles from images acquired by vehicle-mounted cameras will be effective at reducing the cost of maintenance and management because it will reduce the need for inspectors to visit utility poles onsite. In this paper, we propose a convolutional neural network-based image recognition method for classifying the locations of utility poles from roadside images. The proposed method has a unique structure that combines an object detection model that detects utility poles and the areas of public and private land at the pixel level with a relationship recognition model that classifies the locations of utility poles into three categories: public land, private land, or cannot be determined (confirmation required). In the verification results, 242 utility poles were detected with an accuracy of 96.7%, and the locations were categorized with an accuracy of 84.2%, with a particularly high accuracy rate of 91.9% (Recall) for public land. The proposed method can be fully utilized for the primary judgment before the final judgment by the inspector.

The Hanshin Expressway has been conducting the simple repairs of damages found during periodic inspections (hereafter referred to as "the inspector’s first-aid works"). The purpose of it can be divided into two categories: enhancing safety and preventive maintenance. However, a quantitative evaluation of the damage prevention effect through the inspector’s first-aid works for minor damages, which is conducted for the latter purpose, has not yet been performed. Therefore, in this study, we propose a method to quantitatively evaluate the aforementioned effects. Specifically, we employ the Markov deterioration hazard model that takes into account the sample selection bias due to the inspector’s first-aid to estimate the deterioration process. Then, we estimate the number of damages classified under each condition state five years later through simulation. Furthermore, we apply the proposed methodology to inspection data from the Hanshin Expressway to verify its effectiveness.

Material transportation status by crane (loading/ unloading position, transportation distance/ time, and material weight) is considered as one of the indicators that shows the construction process of construction site. The transportation time and material weight can be obtained as numerical data from the chronological record of the weight scale hung on the crane hook. It is necessary to acquire the planar position and transportation distance of the transported object. Photogrammetry technology can be used to acquire these information by analyzing images taken from multiple directions. However, it is often difficult to obtain images from ideal positions and angles due to the obstacles existing at construction sites such as site boundaries, undulations, existing structures, etc.

Therefore, in this paper, we propose a method to obtain the plane position of transported object by using single image taken one direction by focusing on the peculiarities of the transportation mechanism of stationary horizontal jib crane.

In recent years, with the increase in heavy rainfall, the frequency of water-related disasters has been increasing. The Rainfall-Runoff-Inundation (RRI) model plays an important role for predicting flood inundations. Here, parameter optimizations are necessary to achieve accurate predictions. In the future, computationally efficient parameter optimization will be effective for enhancing the accuracy of flood forecasting by accurately estimating parameters that undergo frequent and continuous changes, while minimizing computational load. This study applied the Bayesian optimization for the RRI model to explore an efficient parameter optimization method and estimate the number of iterations required for parameter estimations. The results showed that Bayesian optimization generally reproduced the observed discharge, while reducing the necessary iterations by approximately two to three orders of magnitude compared to conventionallyused optimization methods. Moreover, by examining the posterior probabilities of the evaluation function estimated through Bayesian optimization, we demonstrated that the Bayesian optimization successfully reached to globally optimal parameters without falling into local optima.

From the viewpoint of safety and quality in construction sites, it is not desirable that the motion of multiple workers are not coordinated. Previously, we proposed a method to evaluate the relationship of construction worker’s motion in videos. Specifically, the previous method calculated DTW distance of motion features based on body skeleton positions obtained by utilizing OpenPose. As a result, it became possible to quantitatively grasping whether multiple workers are aware of each other when performing work. On the other hand, the previous method have not yet achieved a quantitative understanding of the subjects and dependencies of coordinated motions. In this paper, it becomes possible to quantitatively understand the subjects and dependencies of coordinated motions by calculating an index that focuses on the time difference of the associated elements for the alignment results obtained by utilizing DTW. It is expeted that evaluation of the relationship between the subject and dependencies of coordinated work by the proposed method will lead to a quantitative understanding of the safety and quality of construction sites.

Distributed Edge-AI is being introduced to the road management approach. To avoid a tight communication network, it is necessary for updating Edge-AI to build a collaborative network of edges that does not assemble all data from each edge. Thus, in this paper, a method of selecting another edge with valid data for updating Edge-AI based on difference in property is proposed. In the proposed method, a similarity score between edges using activation vectors which represent the characteristics of the Edge-AI is calculated. In addition, the edge with the lowest similarity score is selected as containing valid data. By utilizing the proposed method, we expect to update Edge-AI with generality despite using a small amount of data. In the last of this paper, we verify the effectiveness of the proposed method by the experiments of estimating the road surface condition, which is one of the road management approaches.

At construction sites, it is necessary to accurately manage the amount of sediment generated for the effective use of resources. In this study, we proposed a method to estimate the volume of sediment from three-dimensional point cloud data measured using LiDAR, focusing on the sediment loaded in the back of a dump truck during sediment transport, for the purpose of simple sediment volume estimation. First, the characteristics of the LiDAR data were demonstrated through measurements at a construction site. Next, to estimate the volume of sediment in the bed of a dump truck from the three-dimensional point cloud data, the three-dimensional point cloud data of multiple dump trucks were cut into only the dump trucks, meshed, and the volume was calculated. We calculated the sediment volume using the obelisk formula to evaluate the validity of the estimated volume values.

Physical-Informed Neural Networks (PINNs), which directly approximate the partial differential terms of simultaneous partial differential equations, can achieve higher reproducibility than conventional solution methods that use differences. Furthermore, the solution can be obtained much faster. In this study, we investigated the possibility of applying a two-dimensional shallow water equation for calculating the external water inundation of a flood to actual topographical conditions. Since PINNs approximate the partial differential term as a continuous function, they are generally not good at dealing with complex topography with discontinuous shapes. In this research, the water level distribution of complex topography was reproduced by introducing Positional Encoding to improve expressiveness and mitigating discontinuous topography by adding spatial dimension. Furthermore, the calculation time, which used to take about two days, was shortened to a few minutes.

In order to effectively promote the inbound tourism, it is necessary to master the actual tourism actions and concerns of foreign tourists from various countries and regions but has not been enough analyzed. Therefore, we analyzed to use Wi-Fi packet sensor to grasp the actual tourism behavior of foreign tourists in the Yaeyama Islands. Specifically, we conducted a tourism flow survey using Wi-Fi packet sensors to clarify the behavioral characteristics such as visitation rates by country and primary means of transportation and compared them with those of Japanese tourists. To obtain tourist information, we conducted a text analysis of homepages and blogs from Google and Baidu. And we analyzed the frequency of the names of tourist attractions. As a results, while a large amount of web information leads to an increase in the visit rate, the effect is small in island areas, and that air travelers are more likely to be influenced by web information.

CO₂ emissions from commercial and other sectors (offices, commercial facilities, and other buildings) account for about 20% of Japan’s total emissions and reducing them is an urgent issue. In this study, we developed a prediction model for electricity demand and photovoltaic power generation for a building, in order to improve the efficiency of energy management. The model was developed with LSTM, one of deep learning. In this model, following variables are used as explanatory: weather, human flow, periodic components, solar position, etc. The model predicted electricity demand and solar power generation 34 hours into the future with a correlation coefficient of over 0.85 with actual values. We confirmed an annual reduction of over 15% in electricity charges, effective use of storage batteries, the leveling of electricity demand, and effective use for disaster response are possible, through the simulation using the prediction values.

In this study, we verified the accuracy of large language models (ChatGPT) using a database system containing texts in the field of civil engineering. As a result, we confirmed the effectiveness of a extraction task of similar technologies and dialogue task based on the database system operation rules within the validated range. We also confirmed the superiority of prompts based on few-shot-learning in the task of categorizing new technology information. In addition, we confirmed the superiority of prompts created using only job descriptions as personas in the same task.

Evaluating multiple candidate models for prediction based on a large amount of accumulated data and selecting the appropriate model based on the criteria is an important task, which is called model selection. In this study, we have taken up Evidence, Laplace approximation, AIC, and BIC as indices for model selection, and conducted numerical experiments using 100 simulated data sets for polynomial and one-dimensional spatial distributions to investigate the rate at which the correct model is selected. The results showed that the model is selected correctly when the noise level or the second component of the random field is small, but when the noise level or the second component is large, the ratio of correct selection becomes small. Evidence is more stable than AIC and BIC in the numerical experiments. Furthermore, an example of model selection of autocorrelation function using Evidence, AIC and BIC is shown for measured data of geotechnical properties.

This paper proposes a method for automatic detection of dead trees using in-vehicle videos. The proposed method extracts vegetation regions from videos containing various objects based on semantic segmentation. Then, it detects dead trees from the extracted vegetation regions using color information. By presenting the dead tree regions detected by the proposed method to engineers, they can find dead trees efficiently. In the last part of this paper, the effectiveness of the proposed method is verified through experiments using actual in-vehicle videos.

In this paper, we propose a method of acquiring feature representations of record data via the graph neural network to assist in determining the deterioration levels. In the inspection work, multiple deformation images are captured from different angles and distances and stored as record data. However, conventional studies on the deterioration level classification assume the input of a single image for model learning. This makes it difficult to handle the input of record data that has different properties from those of a single image. Therefore, in this paper, to deal with record data, which is a group of multiple images, we construct a graph neural network that can learn the relationship between the individual images and the record data. Therefore, we can acquire feature representations of the record data. In the last part of the paper, the effectiveness of the proposed method is verified through experiments using deformation images obtained during actual inspections.

Visual inspection is important for the maintenance of bridges. However, the decrease in the working population in the construction industry has become an issue in Japan. In addition, visual inspection is time consuming and dangerous in some cases. Therefore, the efficiency, rationalization, and safety of inspection work are required. The current inspection content need decision making based on the experience of the inspector. This can lead to serious accidents due to human mistakes. Inspection methods that utilize AI and UAV can solve these problems. In this study, we performed automatic multi damage detection of UAV images using the Deeplabv3+ model. The problem that UAV images have a large proportion of background and are prone to false positives was improved by background reinforced training. This method is to train a Deeplabv3+ or other semantic segmentation models by standard damage annotated image data, and training it again before use it to real bridge UAV videos by a few background non annotated images to let the background looked familiar to the model. The background reinforced training of UAV images resulted in improved detection accuracy. It is considered that this is because the model learned the characteristics of the bridge and the information around the bridge from the UAV image. After that, by creating a 3D damage model using the image that the damage detection was performed, we proposed a new utilization method of the 3D model.

This study attempts to construct a traffic environment sensing system using a bridge vibration response and proposes a method for detecting traffic using time-domain feature extraction and a neural network. Analysis was performed by examining the feature values of bridge acceleration, the wave group period, and the delay of the waveforms that focus on the mode shape. After evaluating the features, a neural network was constructed to identify the traffic environment around the bridge. According to the verification, the matching rate reached 90 % if there was no congestion, demonstrating the possibility of meeting the accuracy requirement of the mechanical traffic census.

This paper proposes a natural frequency and displacement ratio-based probabilistic damage identification method for bridges using the finite element (FE) model update. When the damage location is known, it can be detected from an appropriate damage-sensitive feature (DSF). However, damaged components are seldom known before inspections. This makes it difficult to find an appropriate DSF and damage identification is sometimes challenging. This paper aims to propose a method to solve this issue by integrating multiple DSFs, natural frequencies and displacement ratio, as a decision-level data fusion approach. They are complementary in terms of sensitivity to damage. In addition, probability density functions (PDFs) of structural parameters are estimated from PDFs of observed DSFs through the FE model update to consider errors and uncertainties in measurement data. An in-house model bridge experiment is carried out to investigate the feasibility. The results demonstrated that the two kinds of damages in a bearing and girder reproduced in the experiment were successfully identified without false positives even when these damages simultaneously occurred.

Recently, many studies have been conducted to develop techniques to detect cracks from concrete surface images using deep learning to improve the efficiency of crack evaluation in concrete structures. However, the annotation of a large amount of training data is required for the construction of a high-performance model, which is a labor-saving issue. In this paper, we propose to apply SimCLR to the training of CNN models to construct highly accurate models with a few annotations. In our experiments in which 5.6% of the training data is annotated with class labels, the results show that our model pre-trained using SimCLR performs better than models pre-trained using ImageNet or other datasets.

The data in the civil engineering field is less data with much variety. Drone river patrols should fly the vast river areas and AI must detect illegal dumping, including general garbage. The patrol drones are not constantly flying and they are rarely captured by aerial images. Thus, it is even more difficult to detect temporary target such as illegal occupation. Previous study has confirmed that the addition of images taken on the ground with different angles of view to the learning data improves learning. However, the number of images is required for training even if the images are taken on the ground. In order to improve the learning of the detection model, this study verified whether the image for data augmentation and the supplement of feature which is less in dataset by generated images by Stable Diffusion.

Social infrastructures in Fukushima Prefecture, including bridges, is affected by natural disasters such as earthquakes and heavy rain. Since the durability and deterioration of an infrastructure are to accelerate under those influences, it is necessary to establish efficient inspection methods. A convolutional neural network (CNN) which is one of the methods of machine learning, the civil engineering filed is also known to as an effective maintenance method in recent year. Previous studies reported learning models as deformation detectors for concrete bridges were developed for each construction and civil engineering offices with using photos taken from road bridge maintenance results in Fukushima Prefecture as training data, and developed models could classified deformations with practical accuracy. Therefore, we tried to improve an accuracy for deformation detections by developing learning models with expansion of training data from maintenance results providing two offices. As a result, we found the improvement of eight points in the classification accuracy of the exposed rebar class and that of four points in the overall accuracy by the learning model with training data expansion.

This study examines the evaluation of landscape images using a deep learning model trained on emotioninducing images. The regression models for pleasant and unpleasant emotions, trained using OASIS and NAPS emotion-inducing images, demonstrated high performance, suggesting their applicability in inferring the quality of landscapes. Inferring the emotional values of landscape images with and without utility poles showed that areas with poles induce displeasure, while areas without them bring about pleasure. Statistical significance was confirmed when comparing these findings with existing landscape evaluations based on fractal dimensions. No correlation was found between the inferred values from the regression model and fractal dimensions, implying the potential to evaluate areas that were difficult to assess using conventional methods, such as color assessment. This suggests the potential for diversification in landscape evaluation.

Stable Diffusion, which synthesizes images by sentences, has been attracting attention. Interpretation techniques for Stable Diffusion have also emerged to indicate the parts of the image that are related to each word in the sentence. In this paper, we propose a method for building an object detection model, which is used re-training, with Stable Diffusion and the interpretation technique. Our proposed method uses Stable Diffusion to synthesize a lot of images of domains similar to the desired object, and automatically annotates the objects in the images with the interpretation technique. Evaluation result shows that re-training with object detection models built by our method resulted in higher detection accuracy than using models trained on the COCO dataset.

Dimensionality reduction methods such as singular value decomposition (SVD) have been widely applied in the earth science field. On the other hand, non-negative matrix factorization (NMF), one of the other dimensionality reduction methods based on matrix decomposition, has been used only to limited applications in the earth science. This study applies the two dimensionality reduction methods, SVD and NMF, to the Radar-AMeDAS precipitation, and compares the extracted features by these decomposition methods. The extracted features are used to reconstruct the original precipitation fields from limited amount of data at AMeDAS observation locations. We compare the two decomposition methods through evaluating errors in reconstructed precipitation fields. For further comparison, this study succeeds in visualizing the features extracted by the two decomposition methods. Through these comparisons, we found that NMF is more robust to reconstruct precipitation fields than SVD. In addition, NMF provided more physically interpretable features than SVD.

The development of text generation models based on Large Language Models (LLMs), such as ChatGPT, has been remarkable. In the field of civil engineering, LLMs are also expected to improve work efficiency. However, since LLMs are mainly trained on documents collected from the Web, there is a concern that they may not be able to generate accurate text due to a lack of training on specialized knowledge in the field of civil engineering. Therefore, as a fundamental study to realize accurate text generation in the civil engineering field, this study attempted to adapt LLM to the civil engineering domain. We proposed an accuracy evaluation method, evaluated the accuracy of text generation in the civil engineering domain using a pre-trained public model of LLM and a model with fine tuning, and discussed the challenges in adapting LLM to the civil engineering domain.

The purpose of this study was to improve the efficiency of construction management operations by centrally managing progress at large-scale construction sites and sharing progress data within the construction site. A prototype was developed to centrally manage three-dimensional point cloud data using a terrestrial laser scanner, a UAV-mounted camera, and a high-altitude camera. Furthermore, a method was devised to visualize changes in geometry using a time axis by assigning time information to three-dimensional point cloud data measured at different times, and a method to visualize construction progress quantitatively by analyzing differences in progress between superimposed three-dimensional point cloud data. Finally, we proposed a use case for an information system. The users are assumed to be the site manager and the contractors at each construction site. The site manager checks the progress status and three-dimensional data of each construction site through the system.

A shuttle bus service for cruise passengers operates between Kanazawa Port and Kanazawa Station at the expense of Ishikawa Prefecture. However, as the number of cruise ship calls increases, there is a financial limit to the number of shuttle buses that can be operated solely at the expense of local governments. In this study, a model of willingness to pay for shuttle bus service using survival analysis was constructed based on the results of a questionnaire survey of cruise ship passengers. Through the analysis in this study, five covariates were estimated: gender, satisfaction with the smoothness of travel to the destination, position on the fee system, number of visits to Kanazawa, and actual use of the shuttle bus. The results also suggest that it is appropriate to set the fare between 100 yen and 500 yen when charging for the shuttle bus.

Due to the impact of the COVID-19, international cruise services were suspended in Japan after March 2020, but resumed in November 2022. In 2023, many cruise ships are scheduled to call at Kanazawa Port, and the city of Kanazawa is expected to be crowded with tourists from cruise trips. One of the characteristics of cruise tourists to Kanazawa is that many of them have a desire to return to Kanazawa again. There have been no studies analyzing cruise tourists in Kanazawa since the deregulation of COVID-19. Therefore, in this study, we used quantification I to clarify the factors that contribute to the willingness of cruise tourists to return to Kanazawa, taking into account their level of concern about COVID-19.

This paper proposes a multi-task classification method that classifies both the distress type and deterioration level at the same time. Conventionally, classifying the deterioration level has been conducted for each distress type by using multiple models. In contrast, the proposed method enables classification of the deterioration level without assigning a distress type to the distress image in advance, by training a single model through loss minimization considering the distress type and deterioration level. In the last part of the paper, it is verified that the proposed method can achieve classification performance equivalent to models constructed for each distress type with a single model by using images of actual distress that have occurred on infrastructure.

In recent years, the number of vacant houses in Japan has been increasing rapidly due to the declining population, and the government has been proactively developing measures to solve the problem of vacant houses and providing subsidies to encourage the expansion of vacant house utilization projects. In Japan, there are many cases of utilization of vacant houses for the purpose of "tourism. On the other hand, there are no organized examples of vacant house utilization throughout Japan, and there is a lack of a general understanding of the status of projects that utilize vacant houses as tourist attractions. In this study, we extract facilities that utilize vacant houses from among tourist spots using travel information websites and conduct a basic tabulation to understand the geographical characteristics of the facilities and the actual status of their users. Non-hierarchical cluster analysis was conducted using the use classification of the extracted facilities and the attributes of the facility users to clarify the use status of the facilities for the purpose of sightseeing. The results of these analyses suggest that vacant houses can be used to create new tourism resources.

In order to estimate the unknown parameters of the antenna model in the simulation of electromagnetic wave propagation for a radar test, we incorporated the ensemble Kalman filter, a data assimilation method, to the finite time domain difference method (FDTD method). As a fundamental study of the application of the method, twin experiment was conducted to simulate radar tests in a simple medium to verify the accuracy of the estimation of antenna model parameters. The validation analysis showed that the proposed method is capable of simultaneously estimating unknown parameters of the antenna model. Sensitivity analysis for various analytical conditions shows that the time increment has a large impact on the ensemble approximation of the error covariance matrix of the state vector and, consequently, on the estimation accuracy. Also, the time increment should be sufficiently smaller than the upper limit of the time increment that satisfies the Courant condition.

Due to depopulation and motorization in rural areas, the number of stores handling daily necessities in small and medium-sized local cities has significantly decreased, resulting in a loss of diversity in store choices for residents when purchasing daily necessities. This study aims to investigate the store selection characteristics of residents living in small and medium-sized local cities located away from urban areas, focusing on their shopping behavior targeting commercial facilities that handle daily necessities. The study utilizes data from KDDI Location Analyzer, which utilizes mobile phone location information, to examine and uncover these characteristics.

Consumer behavior has a wide range of characteristics, such as replenishment of daily necessities, grocery shopping, and whimsical shopping. In addition, purchasing information is analyzed and utilized on a store-by-store basis, so it has been considered difficult to use it in research. However, as the use of big data has been attracting attention in recent years, comprehensive data that is excellent for analyzing purchasing information has become available. In this study, we propose a clustering method that utilizes big data on purchases to characterize consumption behavior in inter-regional travel. This method constructs a dispersion representation of inter-regional travel and intuitively shows the distribution of purchase amounts. The results of this analysis showed that clusters representing characteristics were created and that useful spatial information was obtained for understanding consumption trends in the city.

In recent years, the utilization of AI and data science has been attracted in various engineering fields. This trend is also evident in the field of nondestructive evaluation (NDE), which is essential for structural maintenance. The ultimate goal of NDE is to evaluate the presence of defects inside structures, including their location, and size. Therefore, in the fields of the ultrasonic NDE, the ultrasonic NDE problems reduced to be inverse problems to determine the location and characteristics of defects by analyzing the scattered waveform data obtained from measurement experiments. Academic studies about NDE can be mainly classified into three categories: forward analysis, which involves numerical simulations to understand elastic wave propagation behavior ; inverse analysis, which reconstructs defects and evaluates material constants by utilizing or interpreting the results of forward analysis ; and practical applications of NDE. In this paper, we investigate recent AI and data science trends, as well as their practical applications to these three categories.

Natural disasters such as typhoons and heavy rains have become more frequent and severe in recent years. If the damage of each house in a large-scale disaster can be automatically assessed, it will lead to a reduction in the workload of professional workers such as on-site inspectors and repair person. In this study, we propose a deep learning method using optical satellite images before and after a disaster, focusing on the estimation of damage to individual houses caused by a typhoon. The region used for training is Chiba Prefecture during Typhoon No. 15 (FAXAI) in 2019, and the damage to buildings includes not only roofs, which are relatively easy to detect damage from satellite images, but also other parts of buildings such as exterior walls. From the viewpoint of classifying the degree of damage, the simplest binary classification was considered as the classification task. The multimodal method, which combines images and building attribute information, has a higher recall, i.e., it is harder to miss the damage, than the method that uses only images.

In order to improve the efficiency of infrastructure maintenance and management for port managers, the author has been developing a system for port and harbor facilities using AI. The system targets automatically detection of facility changes of state such as cracks, difference in levels, rust, exposed steel bars. This paper describes the AI-based detection of rust and exposed steel bars in the system. Comparing the past study, I succeeded improvement of detection of rust, but detection accuracies were not sufficient. Further improvement is needed.

In Japan, there have been many cases in which supplies have been in short supply due to large-scale earthquakes such as the Tohoku Pacific Offshore Earthquake and the Kumamoto Earthquake. In such situations, the mental and physical burdens on people in need in times of disaster, such as infants, the elderly, and people with disabilities, are great. The purpose of this study is to understand the disaster awareness of households with infants and toddlers, and the amount of supplies for infants and toddlers in the household. The analysis in this study was conducted using a questionnaire survey, targeting parents of infants and toddlers attending 11 licensed day-care facilities in Kanazawa City. Through this analysis, we clarified the disaster awareness of families with infants and toddlers, and evaluated whether the amount of supplies existing in the home was sufficient for the three-day period during which external assistance was expected to begin arriving.