In the Guidelines for Inspection of Road Earthwork Structures published by Ministry of Land, Infrastructure, Transport and Tourism, the basic method for inspecting deformations is close visual inspection. However, it is also permissible to adopt new inspection technologies if they are deemed reasonable from the perspective of the inspection guidelines, based on collected information on the development trends of new inspection technologies. In this study, we quantitatively analyzed the relationship between various conditions given to PaDiM (Patch Distribution Modeling Framework for Anomaly Detection), a deep learning-based anomaly detection method, for detecting cracks on road reinforced slopes using UAV images as training data and the detection results. The results showed that cracks could be detected by providing UAV images to PaDiM, and we were able to find recommendations regarding the observation methods of road slopes using UAVs.

In this study, we implemented sensor fusion by combining point cloud data and image data obtained from a terrestrial laser scanner, and mapped the segmentation results from the images onto the point clouds. Semantic segmentation of the images was performed using DeepLabv3+ to classify into wheel guard and background. Also, the edge information was updated using the Segment Anything Model, then segmentation information was stored in the point clouds using the camera’s external and internal parameters. Utilizing this information enabled the measurement of bridge widths. By leveraging the detailed information from the image data and the three-dimensional information from the point cloud data, we were able to achieve an analysis that extract detail and structural information, while also efficiently processing large files of point cloud data.

In recent years, there has been a growing trend to utilize non-traditional alternative data in addition to traditional statistical data to evaluate real-time economic conditions. In this paper, we developed a method for estimating the private construction activity in Japan by utilizing information such as construction period and total floor area posted on Notice Signboards of Construction Plan that are placed at construction sites in advance of the construction of mediumand high-rise buildings. We have estimated the private construction activity index from 2016 to 2023 and confirmed that the index shows similar movements to the Indices of Construction Industry Activity previously created by the Ministry of Economy, Trade and Industry (METI) until July 2020. These results suggest that (1) the index presented in this paper could be a useful monthly indicator of construction activity, as it generally accurately captures seasonality and depicts the impact of COVID-19 pandemic as well as the recent downturn in construction activity, and (2) given that signboards are placed before application for building confirmation, the index could be capable of capturing the private construction activity in real time.

It has been important to evaluate pavement structure, and MWD has been developed. MWD can acquire deflection data with short period, but data processing method has not been developed so much. However, it is expected to process MWD data and evaluate road network soundness by machine learning. This research uses Isolation Forest and proposes the evaluation method to detect the section with larger deflection that pavement structure might be damaged. The experiment shows the proposed method can evaluate pavement soundness as the previous methods do, and it is more efficient. The combination of MWD and the proposed method leads more efficient pavement structure evaluation.

Quantitative structural evaluation using measurement data is essential for realizing resilient infrastructure, but the operation of many measuring instruments is a heavy burden. Therefore, in this study, we proposed a method to refine a nonlinear analysis model during an earthquake from measured values at some points using SIS, which is a type of data assimilation. Numerical experiments on a virtual bridge showed that the stiffness of a member, which decreases due to plasticization of the pier base, can be estimated from the acceleration response at one point on the bridge girder with an average error of 21.4%. We also confirmed that the vertical component of the acceleration response is effective as a measured value, and that the rela- tionship between the estimation point and the measurement point affects the estimation accuracy. Further- more, we proposed a method for creating a measurement plan using the proposed method and a flow for assessing the health of bridges during earthquakes.

Utilizing artificial intelligence in the application of geotechnical databases is a prudent approach. It is also important to explore how to effectively train artificial intelligence. A common method is to increase the amount of data. However, collecting data for the target variables to be estimated is not easy. On the other hand, the idea of using text information can contribute to increasing the amount of data. There are “one-hot representations” and “distributed representations” for utilizing text. Therefore, this paper proposes a method to supplement insufficient data by using “distributed representations” for soil types and “one-hot representations” for test conditions (UU，CU，C̅U̅，CD) while using raw data for depth, void ratio, and water content as explanatory variables from the geotechnical database. We attempted to estimate the internal friction angle obtained from triaxial compression tests as the target variable. As a result, a correlation coefficient of 0.956 was achieved.

When inspecting bridges, a countermeasure classification is determined for each part of the bridge or damage. Countermeasure classification is used as basic data when considering repair plans. Therefore, it is important to determine the countermeasure classification based on uniform evaluation criterion. However, countermeasure classification is a qualitative criterion. In addition, the results of countermeasure classification vary among bridge inspectors. Therefore, in this paper, we considered an artificial intelligence algorithm for determining countermeasure classification. The purpose is to support local governments in determining countermeasure classification. We targeted cracks and corrosion of reinforcing bars that occur in small culverts.

To address the recent shortage of engineers, research on rock identification using AI has been progressing. In this study, we investigated the relationship between the number of training images and AI performance by creating and testing eight CNN models with different amounts of training data. The results showed that using approximately 500 images per rock type yielded the highest performance. Additionally, it was found that image generation through data augmentation could cause overfitting. Future prospects include testing models with 500 training images without data augmentation and verifying models with increased rock classification categories.

Filled tofu is considered hygienic and has a low environmental load during production, and its distribution has been rapidly increasing in recent years. The identification of product quality is a time- consuming and labor-intensive task, and there is a need for a technology that can automatically and accurately identify defective products on the production line. The purpose of this study was to classify the coagulation quality of filled tofu using deep learning with color images as input. An imaging system equipped with a polarizing filter to remove specular reflection was constructed, and color images of the film surface of filled tofu were captured on the actual production line. The images were visually labeled into three categories: A, B, and C classes, and classification models were constructed using eight different pre-trained networks. The highest accuracy on the test data was 95.84% with EfficientNet-b4. Visualization of the basis of judgment showed that the model correctly captured the defect features of the tofu surface, with a large weight on the location of bubbles on the tofu surface. These results suggest that color images of the tofu film surface and deep learning can be used to accurately identify the coagulation quality of tofu.

Detailed and quantitative prediction of rail creeping (longitudinal displacement) is important for improving safety against track buckling and reducing management costs. In this paper, a horizontal two-dimensional static elasto-plastic FEM analysis program for rail track is developed. Then, a coupled analysis of this program with an existing rail temperature prediction model is proposed to analytically predict the occurrence of rail creeping and track buckling. The validity of the developed program was confirmed by comparison with an existing one-dimensional analysis model. In addition, it was found that the developed program was able to analyze the behavior before track buckling and the buckling temperature T_A with the same level of accuracy as the existing buckling analysis model. Furthermore, since corner breakage in curves was a problem in the analysis using GIS line data of railway, we proposed a smoothing method which replaces corner breakage with a circular arc of radius R and then performs moving average, and confirmed its effectiveness.

In the construction industry, efficient and effective maintenance and management is required for an in- creasing number of infrastructure structures that are more than 50 years old. Conventional visual inspections have technical limitations and human resources shortages, and new technologies are being introduced. In particular, the development of damage detection technology using AI technology and the utilization of 3D data are being promoted, and technologies such as NeRF, CLIP, and LERF are attracting attention. In this study, we reconstructed 3D space from 2D images using LERF and verified whether specific damage can be detected in 3D space. As a result, it was confirmed that damage detection in the reconstructed 3D space was possible, but the accuracy of damage detection needs to be improved. The accuracy of damage detection in the 3D space could be improved by additionally learning damage in CLIP.

Damage diagnosis for road structures in Japan is carried out once every five years, and is based not only on an objective evaluation of the damage conditions, but also on a comprehensive evaluation that incorporate the expert opinions of inspecting engineers. On the other hand, much of the research into using AI to improve the efficiency of inspection and maintenance has focused on objective facts, and it is possible to perform comprehensive damage diagnosis by using the expert opinions of inspecting engineers. Therefore, for language-based information, it is necessary to understand the complex relationships between rich expressions through the modeling of the relationship between images and text. In this paper, the bridge damage diagnosis with CLIP was investigated, using the findings and damage images recorded in xROAD and aiming at the inherit language-based information. Specifically, features were extracted from each of the findings and damage photographs listed in xROAD, and a VQA model for classification data was built by learning a feature vector that integrates these features. The results suggest that by using language-based information, such as the inspector’s findings, CLIP-based bridge damage diagnosis can improve generalization performance for deep learning-based tasks.

In Japan, transportation infrastructure such as roads and bridges, which forms the core of public infra- structure facilities, was intensively developed during the period of rapid economic growth in the 1960s. However, more than 50 years have passed since their construction, and they are aging and deteriorating along with the rapid development of motorization. The number of such infrastructure facilities of deterioration is increasing with each passing year, making efficient maintenance and management of aging infrastructure an urgent task. In order to carry out efficient maintenance management, it is necessary to develop a method to accurately and quickly evaluate the degree of deterioration of structures. In this study, we focused on road bridge slabs, which are a major transportation infrastructure structure, and showed that it is possible to infer the degree of deterioration of the slabs by applying a convolutional neural network method to the strain waveforms of the slabs caused by traffic loads traveling on the bridge surface.

Much of the infrastructure built during the period of high economic growth is now aging, and issues such as a shortage of engineers and long working hours have become significant problems. As a result, there is a growing demand for more efficient and automated inspections. Currently, bridge inspections are largely dependent on the manual labor of engineers, which can be considered inefficient. In this study, we developed a method targeting long-span bridge with a wide inspection range. Using a permanently installed inspection vehicle, we photographed the underside of the stiffening girders and employed image diagnosis technology to automatically measure the location and area of corrosion. By reflecting these results in a 3D model known as BIM (Building Information Modeling), we can achieve more efficient management. Furthermore, we verified the effectiveness of this method on an actual long-span bridge.

The bender element (BE) method is a low-cost, nondestructive method for determining the shear modulus of soils. Therefore, methods have been proposed to determine the arrival point of S-waves, which is often difficult in the BE method, but there is no general-purpose method for all types of soils and test conditions. In this report, a deep learning model using high-dimensional features in the frequency domain was developed and validated for general-purpose and highly accurate S-wave arrival point prediction. The model was trained using 4/5 of tens of thousands of artificial waveforms, and a model with high validation accuracy was created for the remaining 1/5 of the artificial waveforms. The best prediction error for the 173 experimental data used to test the model was 11.88%, indicating a certain level of significance of the high-dimen-sional features. On the other hand, the prediction accuracy was lower than that of Momiama and Ogino’s previous model, again demonstrating the usefulness of the low-dimensional features.

Critical Line (CL) for alerting landslide warning information in Japan is set based on the past landslide disaster and rainfall information. The accuracy of alerting warning can be gained by re-set the CL considering terrain, geological and vegetation information. This paper presents a way of estimating thresholds of CL (60-minute total rainfall and soil water index (SWI)) from rainfall, terrain, geological and vegetation information by using three types of machine learning algorithms (Random Forest (RF), XGBoost (XGB) and LightGBM (LGB)). Accuracy of the models established by these algorithms were evaluated, and important feature values which gains the accuracy of the model were clarified by evaluating feature importance and Recursive Feature Elimination (RFE). Basis for the estimation of the models were explained by using SHapley Additive exPlanations (SHAP) which can clarify contribution of each information to the results of estimation. Following results were obtained from this study. 1) For estimating thresholds of 60-minute total rainfall, using RF with extracted 10 explanatory variables from 100 variables was the highest coefficient of determination (R²) for testing data among the other models. For estimating thresholds of SWI, using LGB with extracted 10 explanatory variables was the highest R² for testing data among the other models. The extracted variables include geological type number which is distributed by Ministry of Land, Infrastructure, Transport and Tourism in Japan as a geological feature. 2) In the SHAP estimation of 60-minute total rainfall, all models showed a clear tendency for larger values of the maximum 24-hour rainfall, June and August rainfall anomalies to contribute to an increase in the 60-minute total rainfall. For the SWI, the models with the highest R² showed a clear tendency for larger values of 12-hour rainfall and the August rainfall anomalies to contribute to an increase in the SWI.

After an earthquake, it is crucial to judge the necessity of traffic regulation and implement emergency measures based on load-bearing capacity and drivability to facilitate early road clearance. The diagnose on road bridges has traditionally relied on past earthquake damage and recovery response record as a reference, however, inconsistencies in diagnostic results have become an issue. Additionally, considering the shortage of engineers during disasters and the need of lightening their workload, it is urgent to develop emergency inspection methods that enable them to diagnose remotely. Therefore, in this study, we examined a remote diagnostic system using digital twins as a new emergency inspection method. We also organized the benefits of seismic emergency inspections using digital twins and identified the challenges that need to be addressed to implement this approach.

In straddle-type monorail, the expansion joints installed at boundary of the track girders can experience cracks and spalling in the cover concrete due to the train induced impact loads. The engineers conduct hammering inspections four times a year, requiring a huge amount of manpower. In this study, to automate the damage detection in hammering inspections of expansion joints, the modal characteristics are identified by multi-point hammering and reciprocity theorem, and the damage effects on the modal characteristics are investigated. As a result, the modal amplitude of the claws at damaged concrete increases around 1100 Hz. Utilizing this characteristic, a damage evaluation and a damage area estimation method based on hammering the four claws of the expansion joint plate are proposed.

Since some studies and implementations of UAV bridge inspections to efficiently recognize structural damages are recently active, it is required to achieve not only effective cost reductions but also inspection performances of required damage recognition accuracies. In this study, a flight path optimization for the inspection of a steel plate-girder bridge was verified by the flight simulation in a 3D model of the bridge in the virtual space. Based on surveying the bridge inspection standard, damage-prone areas were identified, and the view point of interest (VPI) and the viewpoint importance levels were determined. Then, the optimal path was calculated using some optimization methods. Comparing the optimized flight paths between without and with considerations to the viewpoint importance, ant colony optimization (ACO) method showed good performances to derive the shortest path length even in the viewpoint importance was considered. Although the path length became longer when the viewpoint importance was considered, and the flight simulation results could confirm that the same angle of views to bridge members required to be visually inspected at a short distance were obtained.

Investigating appropriate investment policy for flood risk reduction has become important considering the increased heavy rain disasters due to climate change. This study proposes using the reinforcement learning to minimize the summation of investment and disaster damage costs. We applied the reinforcement learning for a mathematical risk model of flood damage, which is computed by 100-year projected precipitation changes. From the simulations using the trained results, investment plans suggested by the reinforcement learning successfully reduced long-term total costs by 40 % compared to straightforward investment plans. It has been suggested that investment plans of reinforcement learning tend to invest in disaster prevention prior to the rise of flood risk, so that the disaster damage costs can be reduced.

With the spread of 3D-based technologies in river surveying, design, and evaluation, it is now possible to simultaneously consider flood control, the environment, landscapes, and utilities. This will bring solutions for issues in conventional river design, where the environmental aspect is often set as secondary to flood control measures. We conducted a case study to examine the practical efficacy and issues involved with using 3D technologies in a river holding a couple of fine-resolution topographic datasets. Through our study, we identified that they are helpful for building a reliable model set and supporting the introduction of more realistic calculation conditions, such as spatial grain size distribution. However, this approach brings not only better reproducibility but also additional concerns, such as the need to adjust riffle-pool segmentation parameters.

In this study, we examined the effect of different prompts on detection accuracy with the aim of improving the accuracy of image classification using the Large Vision Language Model (LVLM) in the civil engineering and construction fields. The results suggest the effectiveness of using prompts that represent the objects to be detected by showing specific examples in the validation range. The results of this study suggest the effectiveness of using prompts that represent the objects to be detected by showing specific examples in the validation area.

A convolutional neural network (CNN) has alos been gotten attention in the field of civil engineering in recent years, which is demonstrated useful inspection methods. A previous study has been reported in which the learning model of corrosion detector was developed using photographs of road bridge inspection results in Fukushima Prefecture as training data. However, the further expansion of training data has needed to improve the classification accuracy. In order to improve the accuracy of corrosion detection compared to previous studies, this study was focused on brightness changes of the preprocess in training data rather than geometric changes on images such as a rotation and flipping, which a common data augmentation method in the CNN. The training data was used the photographs of the condition of road bridge inspections in Fukushima Prefecture, which was preprocessed 90º rotation, left-right flipping and also contrast reduction and enhancement, and histogram flattening. The classifier for the corrosion detection was developed by applying those training data. As a result, we found that the classification accuracy of corrosion class could be maximized using the learning model trained data of contrast reduction preprocessed.

The development of large language models has been remarkable, with models comparable to OpenAI's GPT-4 being announced by various companies. Additionally, the trend towards multimodal capabilities is progressing, with models like GPT-4V(ision) that handle image input and GPT-4o with enhanced image comprehension. This paper reviews the technical trends of large language models and examines their use cases in the medical and civil engineering fields. Furthermore, as a method to incorporate specialized knowledge in civil engineering, we have organized the mechanisms and challenges of RAG, which is currently considered to be cost-effective. Based on the implementation results of RAG, we have identified the challenges of how to integrate domain-specific knowledge into workflows as a knowledge platform.

This paper shows the usefulness of the data augmentation method that modifies the frequency characteristics of sounds in the problem of detecting wood broken sounds occurring in large earthquakes. In this study, a CNN model is used to classify the wood broken sounds and the other environmental sounds. The CNN model is trained with the spectrograms that are converted from sound of 1 second. First, it is investigated whether if the wood broken sounds are correctly classified or not even though the recording devices or environments are different between the training and the testing. The result shows that it is difficult to classify the wood broken sounds recorded with different devices. Next, the frequency characteristics of sounds are modified to simulate the sounds recorded with other types of devices. As a result, it is confirmed that the accuracy of classifying the wood broken sounds is improved by training the simulated sounds.

The purpose of this study is to confirm the applicability of the Local Outlier Factor (LOF) method in non-destructive inspection using impact sounds to detect damage caused by projectile impacts on RC spec- imen coated with polyurea resin, aimed at preventing delamination and the scattering of scabbing fragments. In experiments targeting RC specimen damaged by projectile impacts, the impact sound spectra during hammer strikes on the resin-coated layer are used as input data for LOF-based damage assessment. The study confirms the feasibility of damage detection using LOF and examines the influence of the number and acquisition position of training data on the assessment results.

This paper examines the progression of soundness for bridges managed by Fukui Prefecture that have early deteriorated due to chloride attack, alkali-silica reaction (ASR), frost damage. As a result, many of the bridges with deteriorated ASR showed a decline in health. To efficiently manage these bridges in the future, machine learning was used to analyze the factors that influence the prediction of the deterioration of these bridges. Consequently, it was possible to identify explanatory variables that influence the prediction of each deterioration, such as geographical conditions and environmental conditions. In particular, the distance to major rivers was considered as an explanatory variable for ASR deterioration, since it has been suggested that ASR deterioration is related to the rivers that supply reactive aggregate.

Recently, concrete structures are required to be properly maintained. However, inspections of concrete structures are labor-intensive. The purpose of this research is to improve the efficiency and accuracy of the soundness evaluation of tunnel linings. Until now, we have been investigating the method to detect defects such as cracks on tunnel linings using image analysis. In addition, we also have been developing the method to identify the three types of critical defects from those defects. We have confirmed that these methods can be used to accurately detect the three types of critical defects in images of tunnel linings. The results of this research may contribute to efficient and accurate the soundness evaluation of tunnel linings.

In the inspection of road bridges after a major earthquake, it is often difficult to deploy bridge inspection vehicles or to set up scaffolding due to lack of materials and equipment, and inspection of bearings is often forced under such conditions. Therefore, a practical study was conducted with the use of UAVs and selfie sticks as inspection methods in the absence of scaffolding, with the aim of reliably capturing deformations using cameras attached to these vehicles. As a result, the following findings were obtained: (1) hands-on inspection was sufficient for small single-diameter bridges with easy access, (2) the selfie stick was effective for many slab and girder bridges, and (3) UAVs were effective for superstructures with high girder heights.

For the maintenance of civil engineering structures, it is important to accumulate records of damage and deterioration as materials for diagnosis and the next inspection, to determine whether the location and appearance of the damage is progressing, and whether repairs are necessary. For bridges and damage, videos were used as input data, and 3D scenes were constructed using NeRF (K-Planes), and a 3D solid model was constructed from the NeRF information, and a study was conducted on the use of this for digital twin inspection. A NeRF model was also demonstrated outdoors, where actual infrastructure inspections are carried out, to confirm the accuracy of the 3D model constructed and identify issues unique to field inspections.

In this study, storm surge prediction in the Ariake Sea using a hierarchical neural network was conducted. The network was trained with only typhoon information (position, central pressure, and maximum wind speed radius) as input, using a dataset generated from storm surge calculations performed by FVCOM and d4PDF. As a result, we confirmed that storm surge could be predicted with high accuracy even with only typhoon information. Furthermore, we found that inputting the typhoon location as a distance angle from the prediction point, instead of latitude and longitude, improved the prediction accuracy even when using the same typhoon information because of improving the interpretability of the feature values. We evaluated the prediction accuracy using actual typhoon data as input and presented issues for the future operation of storm surge prediction.

In this study, we examined methods to make the AI decision process interpretable and improve performance in the detection of corrosion and damage in tunnel lighting fixtures. Specifically, we utilized ResNet18 to extract features from images of the lighting fixtures and selected these features through decision tree analysis. Subsequently, we applied Grad-CAM to visualize which parts of the selected features the decision tree focused on, thereby clarifying the AI's decision-making process. Furthermore, we focused on the importance of distinguishing between classes that require maintenance actions and those that do not and aimed to improve the accuracy of classes with poor classification performance. Specifically, we used a random forest to explore alternative features and constructed a new decision tree. As a result, the model's performance was enhanced, particularly in improving the classification accuracy of Class 2, thereby leading to a model better tailored for determining the necessity of maintenance actions.

The purpose of this study is to examine the applicability of a multi-class classification method for rebar corrosion using impact sounds based on a neural network model. In addition, the Influence of the composition of the training dataset (the number of data, contamination of mislabeled data) and the number of intermediate layer nodes on classification results were investigated. As a result, it was confirmed that by using a neural network model with three output nodes, it is possible to classify rebar corrosion into three classes: corrosion level of 0% (no corrosion), corrosion rate 1% and corrosion level of 6%. Furthermore, it was found that in order to improve classification accuracy, it is desirable to collect as much training data as possible. On the other hand, in the case that the number of intermediate layer nodes was excessive or in the case that mislabeled data was contaminated into the training dataset, the tendency for classification accuracy to decrease was confirmed.

In the maintenance of civil infrastructures, the importance of efficiently evaluating the load bearing per- formance through numerical analysis is increasing. In this study, we proposed a method to efficiently eval- uate the load bearing performance of civil infrastructures using point clouds, which have become increas- ingly measured and accumulated in recent years. A deep learning network was specifically designed to estimate load bearing performance indices from geometrical features of point clouds. The network was trained using a virtual point cloud dataset of a simple beam of H-steel. We proceeded to estimate the mo- ment of inertia and yield load as the load bearing performance to compare the estimated results with the true values calculated from the actual size. The results show that the moment of inertia and yield load can be estimated with an error of approximately 20% of the true values, and this method is applicability to estimate load bearing performance.

Physics-Informed Neural Networks (PINNs) that directly approximate the partial differential terms of coupled partial differential equations are a promising method for analyzing physical phenomena more quickly and with higher data reproducibility, potentially reducing computation times from as long as two days to just a few minutes. However, there are trade-offs between accuracy and speed due to (a) the de- pendence of error convergence on the strictness of physical laws, where ensuring the strictness of physical laws incurs greater computational loads-an important issue for the analysis of external water flooding in complex terrains; and (b) the inherent difficulty with clear terrain conditions since the entire spatiotemporal domain is represented as a continuous function, leaving many challenges in applying PINNs to real-world terrain problems.

In this study, by improving and introducing a hierarchical grid-based position embedding method called K-Plane into PINNs, we demonstrate that (a) it is possible to maintain high fidelity to physical laws while reducing the number of sampling points for calculating PINN losses, thus enabling faster computations; and (b) by pre-training K-Plane with terrain conditions, it is possible to represent water surfaces in accord- ance with natural terrain conditions, such as inside and outside levees. The feasibility of applying this method under real terrain conditions was confirmed. With terrain conditions and observational results, sim- ilar computations can be realized anywhere.

Fruits and vegetables are subject to loss and waste at each stage of the food supply chain from harvest to consumption, and there is a need for technologies that contribute to loss reduction in the post-harvest food supply chain. The objective of this study was to build a model to predict the rate of weight loss using color and fluorescent images as input. Color and 365 nm excitation fluorescence images were captured and excitation emission matrix (EEM) were measured over time in cherry tomatoes of different fruit colors. Three models were constructed with red cherry tomatoes, yellow cherry tomatoes, and both red and yellow cherry tomatoes as input. As a result, RMSE, MAE, and R² for the model with input of multicolor cherry tomato images were 0.853, 0.676, and 0.660, respectively. This result was almost as accurate as the single color model. It was suggested that the rate of weight loss of cherry tomatoes of different fruit colors could be predicted using a single model.

To prevent road cave-ins in advance, we constructed a supervised deep learning model (Multi-Layer Perceptron: MLP) using 9 variables and applied it to the real field to assess the risk of road cave-ins. This deep learning model had issues with overfitting, so we tried to resolve the problem by properly controlling the number of iterations after determining the optimal hyperparameters. Consequently, we were able to mitigate overfitting with a slight decrease in accuracy of less than 3% compared to the evaluation metrics during training and reduce the false-negative error of road cave-ins by up to 17%. Moreover, excluding variables of categorical and different attribute information did not significantly affect the accuracy, sug- gesting the potential for efficiency in collecting input information. In the future, it is desirable to update and accumulate supervised data to construct a more accurate prediction model.

Intersection geometry diagram including road markings are important data for advancing the digital transformation (DX) of road administration. There is a need to establish efficient data collection and generation methods. Previously, a method of generating diagrams using semantic segmentation (SS) based on aerial photographs has been proposed by the authors, but the challenge has been remained that the accuracy in areas with occlusions such as pedestrian bridges declines. To overcome the defect of the previous methodology, this study proposes a method that combines 3D point cloud data and aerial photographs to perform semantic segmentation considering occlusions. This method identifies occlusion objects from the 3D point cloud data, projects their latitude and longitude information onto the aerial photographs, masks the corresponding pixels, and performs SS inference. As a result of the verification, it was clarified that it is possible to identify the components of intersections around occlusions accurately, and that there is an appropriate point cloud density that maximizes identification accuracy.

This study created training data by manually classifying road conditions into "snow-covered" and "non- snow-covered" categories using images from road monitoring cameras installed on major roads, serving as a support tool for road administrators and residents to quickly check road surface snow conditions. Additionally, an AI model for automatic road surface snow condition detection was developed and tested using 400 images taken during both day and night. As a result, the accuracy rate was approximately 85% or higher.

In periodic inspections of bridges, damage conditions are assessed through close visual inspection, followed by evaluations of damage extent and structural integrity. The use of deep learning models is expected to streamline the assessment of damage. Constructing high-performance deep learning models requires extensive data collection and annotation with class labels tailored to specific tasks. This article applies five representative anomaly detection models (Ganomaly, PaDiM, PatchCore, FastFlow, and EfficientAD) to the problems of crack detection and segmentation in concrete structures, proposing a method to build models using only normal data. Evaluation experiments assess the effectiveness of each model in crack detection and segmentation. Additionally, this article shows limitations of anomaly detection models and discusses challenges towards practical applications.

In recent years, the release of various data platforms has made it easier to conduct analysis of infrastructure facilities in multiple regions. In this study, we analyzed bridge inspection data accumulated in bridge maintenance management systems in multiple regions, organized the deterioration trends, and examined risk assessment indices that can be used universally by each road administrator, considering the extent and progress of deterioration. By evaluating the deterioration state of a group of bridges using risk indicators, we summarized the deterioration trends common to all regions and the differences in deterioration trends for each health assessment category, and examined use cases for utilizing risk indicators for diagnosis.

In recent years, data platforms such as xROAD have been increasing. Combining these data enables various cross-regional analyses, but analysis using all the data requires ample computing resources, and duplication of data leads to waste in the allocation of analysis resources. In order to solve these problems, this study examined a method to construct a bridge deterioration estimation model by extracting the minimum necessary for the analysis from a large number of explanatory variables. As a result, we confirmed that the accuracy of the model using the explanatory variables narrowed down by our method as teacher data retains some accuracy when all variables are used.

There is an increasing number of cases where dam operations are optimized using deep reinforcement learning based on meteorological conditions and various dam quantities. However, in making decisions on dam operations, dam discharge operations are judged based on various circumstances such as stakeholders in the dam basin and CCTV camera images, in addition to meteorological conditions and dam quantities. It is difficult to model these values of dam discharge operations as reward functions in deep reinforcement learning. Recently, large language models (LLMs) have been able to implement deep reinforcement learning based on human values through Reinforcement Learning from Human Feedback (RLHF), achieving more accurate responses. In this study, we applied RLHF to a dam discharge operation model using deep reinforcement learning and constructed a dam discharge operation model that incorporates human values.

In our country, AI-based survey methods are being promoted to streamline traffic volume surveys. Existing research has shown that vehicle section identification using AI can count sectional traffic volumes by vehicle type, but accuracy decreases due to flares and scenery reflections. Additionally, vehicle types can be determined by recognizing the leading number of the classification code on license plates, but the recognition accuracy decreases when the characters are unclear. Therefore, this study proposes a method to determine vehicle types by recognizing classification codes when the characters are clear and by using vehicle section identification results when the characters are unclear. Applying the proposed method to videos taken at three different locations resulted in a high accuracy with an F-measure of 0.95 or higher at all locations. In the future, we aim for early practical application in automotive traffic volume surveys by accommodating license plates with designs.

In this study, we propose a method to estimate the position of a pedestrian using only the vibrations generated during walking. The vibrations in the floor generated during walking attenuation with distance. Therefore, we focus on the temporal changes in the intensity of each frequency to estimate the distance. With a large amount of training data, the characteristics and environment of the vibration source can be learned, enabling the estimation of previously difficult-to-detect vibration sources. Enter the STFT image generated from the vibration data obtained from the sensor array into the CNN to output the estimated distance. In the experiment, the distance was estimated using multiple piezoelectric elements, and the estimation accuracy of the pedestrian’s position was evaluated. The results showed that the average error rate between the actual stepping points and the estimated human positions was 69.7%. Verification using walking data confirmed that the average error rate was reduced to 36.7%.

In Japan, the number of bus users is decreasing due to the nationwide increase in automobile use, population decrease in rural areas, and outflow of population to the three major metropolitan areas. In response, bus companies are planning efficient bus routes by leveraging data such as the number of passengers. Thus, mechanical passenger counting technology has garnered attention for improving work efficiency and reducing labor, and various methods have been proposed for installing such devices inside buses. However, if facilities and organizations other than bus companies wish to use these data and onboard counting methods, they must pay significant financial charges to the bus companies. Therefore, in this study, we developed a deep-learning-based method for counting the number of passengers disembarking from buses using video camera images acquired at bus stops. Through demonstration experiments, we highlighted the usefulness of the proposed method and identified areas for improvement.

In this study, a method using a surrogate model based on machine learning was considered to propose a fast and accurate estimation method for the maximum displacement of RC beams subjected to impact action. Six representative estimation models were used: generalized linear model (GLM), deep learning (DL), decision tree (DT), random forest (RF), gradient boosting decision tree (GBDT), and support vector machine (SVM). As a result, the applicability of machine learning to the estimation of the maximum displacement of RC beams subjected to impact load was confirmed. However, more detailed consideration is required by identifying the range of explanatory variables in the training data and outliers in the experimental data. Within the scope of this study, it was revealed that when using RF in the case that selected three explanatory variables are used, it is possible to estimate the plasticity ratio up to about 12 with a certain degree of accuracy.

In recent years, the importance of non-destructive testing has increased due to the aging of civil engineering structures and nuclear equipment. Ultrasonic non-destructive testing is one of the most widely used non-destructive testing methods and is used extensively from the manufacturing processes of various materials to the on-site inspections. Recently, advanced ultrasonic measurement techniques, such as Laser Ultrasonic Visualization Testing (LUVT), have been developed, and attempts to use machine learning for automatic inspection are also being explored. However, the lack of anomalous data with defects poses a barrier to improving the accuracy of automated inspection through machine learning. Therefore, in this study, we propose a method for automated LUVT inspection using an anomaly detection approach with a diffusion model that can be trained solely on negative examples (defect-free data). We experimentally confirmed that our proposed method improves defect detection and localization compared to general object detection algorithms used previously.

The large rainfall ovserved in Murakami City, Niigata Prefecture, on August 3-4, 2022, caused simultaneous slope failures on many slopes and debris flows were observed in the valley. Detailed observation of satellite and aerial photographs revealed that the collapse sites were unevenly distributed at the ridge boundaries. In order to find the causes, we analyzed the latent failure characteristics of the slope failures using transfer learning and the triggering factors based on the actual observed wind direction and the analyzed rainfall for each rainfall mesh from the beginning of the rainfall. The results suggest that the observed slope failures were caused by localized rainfall on the wind-back slope due to the orographic precipitation.

There has been considerable debate about the necessity of removing utility poles, yet there are significant challenges in determining the most appropriate priorities. Furthermore, while the removal of utility poles has been discussed from the perspectives of scenery, traffic safety, and disaster prevention, there has been limited research from the perspective of disaster prevention. With regard transportation roads from the perspective of disaster prevention, discussion has also been limited to remove utility poles from emergency transportation roads from the perspective of removing obstacles to disaster recovery.

This study, focused on the urban area of Ishigaki city, which is at risk of tsunami disasters. The priority of removing utility poles from the perspective of their importance as evacuation routes was examined based on an analysis of the evacuation behavior of residents and tourists. Consequently, the number of individuals evacuating a given road can be utilizes to identify which roads should be prioritized for construction without utility poles.