The civil engineering and construction fields are facing the problems of a declining population, a high aging rate compared to other industries, and a shortage of workers, including construction equipment oper- ators, and there is an urgent need to achieve automated construction and automated operation of construc- tion equipment by using AI. This paper outlines the author's recent approaches to automating excavation and ground leveling tasks in earthwork operations using imitation learning and reinforcement learning, and presents the results of experimental validation for excavation and grading tasks in simulations and simple experimental environments.

As crop predictions can assist in decision-making regarding crop selection, time period, duration, etc., there is a need for an automatic and accurate methodology for predicting yields in large agricultural areas. This study aimed to determine the pattern of distribution of yields using the videos of sugar beets captured from a camera attached to beet harvester. Deep learning networks for object detection, such as YOLOv2, YOLOv3, and EfficientDet, were employed to use the videos to automatically detect sugar beets on the conveyor belt of the harvester. The mean average precision of sugar beet detection with these detectors was over 0.95, indicating the high accuracy of object detection. Further, identical sugar beets in consecutive video frames were identified by object tracking using Kalman filter. This aided in the accurate counting of sugar beets flowing on the conveyor belt. The counting was performed with an F-value of over 0.95. For example, 840 of 847 sugar beets were properly counted. Also, the weight of each sugar beet detected by the deep learning network was estimated using regression technique, based on its long and short axes. The correlation between the size and weight of sugar beets was computed prior to the experiment on site, which enabled to predict the total weight of the harvested beets. The error of yield prediction using YOLOv3 and EfficientDet was about 3%. Based on the predicted yields of sugar beet that was harvested, the geospatial information was associated to each of the detected sugar beet as the camera recorded the GPS information while recording videos. Finally, the distribution pattern of sugar beet yield was visualized and the variations were analyzed.

When designing a soil structure, results obtained from a soil investigation are converted into parameters that are directly needed for design calculation; a conversion error in such a case, however, is considered as a challenge to be addressed. If a soil constant necessary for numerical analysis of soil can be estimated accurately by using AI, the quantity of information required for analysis is expected to be increased. In this study, Internal friction angle data obtained by triaxial compression tests were arranged from KANSAI Soil databases and a trial estimation was made on values for Internal friction angle by means of artificial neural networks. First, 504 data sets collected from the Kansai area soil information databases (data of Kobe City) were created. As a result of the above, the coefficient of determination became 0.657, showing a Internal friction angle by excluding the UU test. Regarding future prospects, increase in AI data, comparison of AI algorithms, estimation of other soil constants, and verification of applicability per region are scheduled to be conducted.

Rainfall-induced slope failures can cause road closures and a significant delay in the recovery process. Given the increase in the frequency of extreme rainfall events, the prioritization of countermeasures against slope failures is needed to achieve the disaster mitigation. In this paper, the methodology for determining the prioritization of countermeasures against slope failures due to heavy rainfall using machine learning and probabilistic rainfall intensity is presented. The rainfall index calculated based on long- and short-term effective rainfalls is used as one of the explanatory variables representing a rainfall intensity measure considering the effects of time variation of rainfall on the occurrence of slope failure. The training data used in the machine learning are re-sampled to avoid overfitting caused by imbalanced class samples. Random Forest and LightGBM are used to develop the prediction model of rainfall-induced slope failure. As a result, LightGBM shows better performance than Random Forest. In addition, the probabilistic rainfall intensity is defined as the rainfall index corresponding to a return period and estimated using the generalized extreme value distribution. The locations where slope failures can occur are evaluated using the prediction model based on LightGBM and the probabilistic rainfall intensity. A illustrative example demonstrates that the proposed methodology can be used to determine the countermeasure prioritization based on return period.

Some recent studies have been conducted to detect the deterioration status of concrete structures based on AI technologies by hammering sounds. The focus of this study is to examine the effects of hitting conditions of training data and test data used in judgment for re-bar corrosion cracks using AI technology on the judgment results. For this purpose, a hammering sounds test with the RC specimens was conducted to determine the corrosion cracks by using a neural network, the effects of the difference and the range of variation in maximum impact force on the judgment results were considered. As a consequence, the effects were shown that the harder the maximum impact force, the better the accuracy when the maximum impact force at the time of hammering is coincident for training data and test data. In addition, it was confirmed that the accuracy tends to decrease in the case of the maximum impact force of both data is discrepant and in the case of the range of variation in the maximum impact force is wide. It hence has become clear that it is desirable to be harder the maximum impact force, to narrow the range of variation, and to make the maximum impact force between training data and test data approximately coincident at the time of recording hammering sounds.

In the infrared thermography method, which remotely detects inner defects by capturing thermal images of concrete, damaged areas are often overlooked by human judgment. Although there is a movement to introduce CNN-based automatic detection to the infrared method, sufficient accuracy has not been obtained due to the lack of training data. Therefore, in this study, we focus on self-supervised learning. Self-supervised learning has the potential to achieve high accuracy with fewer teacher labels. In this study, we present an example of how to introduce self-supervised learning to the infrared thermography method and verify its effectiveness.

Physics-informed neural networks (PINNs) have developed as a deep learning method that can output solutions based on physical laws. This study verified the applicability of PINNs to unsaturated seepage simulation through the reproduction of test results using soil column tests, and discussed the methods to construct an accurate PINNs model. It was clarified that the unsaturated seepage simulation using PINNs was sufficiently available. Furthermore, we could construct an accurate PINNs model for unsaturated seepage simulation by using the sum of squared error as the loss function and the pressure head as the physical quantity output from the PINNs model.

With the need to establish efficient methods for infrastructure management, it is important to build a system that anticipates the social implementation of technologies such as deep learning. In this study, a web system for bridge inspection applications utilizing image captioning technology is developed, and the optimal design of the system design obtained in the process is summarized. Furthermore, the system developed in this study is designed to take advantage of data obtained in the process of the system being used by users, so that the system can continuously incorporate civil engineering expertise. We also showed that adding new data from users' use to the teacher data leads to an improvement in the accuracy of deep leaning.

Generalization performance of a support vector regression model for the bender element test of which the S-wave arrival point is often difficult to determine on the received signal has been examined for the purpose of assisting experimenters' determination. Based on linear theory, 8960 received signals were numerically generated using realistic parameters. Then the model was trained with predictors of 11 dimensions which were extracted from the signal shape and the test condition as well as the true arrival point of the S-wave. The model has been validated by comparing the predicted values with the true arrival points using the synthetic signals. The prediction has also been made for a series of received signals in actual bender element tests which were conducted in Toyoura sand, a legorith simulant, and a peat and compared with arrival points which were determined by an expert. The prediction has substantially agreed with the expert's determination.

This study proposes how to improve the prediction accuracy of deep neural network (DNN) using the feature extraction of frequencies from input data in the preprocessing of the DNN simulation. This method requires the Fourier transform and inverse formation to extract frequency-based features from the data. The features are added into the input data as new data. The method was applied to the water level prediction at a pond nearby a drainage pumping station. The transform technique extracted daily- and half-day- frequencies for the dataset collected at a field for about eight years. This study compared the prediction with both frequency features with that without the features. The results show that the prediction with the proposed method was improved by 3-5% in a root mean square error (RMSE) for 1-6-hour lead time (LT) in a comparison of the conventional method . The future prediction with an extra dataset was also improved up to 4% in RMSE.

In recent years, 3D point cloud surveying instruments such as Unmanned Aerial Vehicle and Terrestrial Laser Scanner have been widely used in the field of civil engineering and construction. However, the use of such surveying instruments is not widely used in the construction industry, especially for small-scale earthworks, due to their high operating costs. Under such circumstances, mobile terminals equipped with LiDAR have been released. and many applications that enable the acquisition of 3D point cloud data have been released. However, there are few applications that support public coordinates, which are necessary in the field of civil engineering and construction. Therefore, this paper aims to convert 3D point clouds with arbitrary coordinates into public coordinates with much greater ease than conventional methods. The goal was also to ensure that the results of converting the coordinates of surveyed point clouds of actual smallscale earthwork sites using the results of this research fall within the measurement accuracy specified by as-built management guidelines of the Ministry of Land, Infrastructure, Transport and Tourism.

In order to improve the safety of railways, it is indispensable to conduct a detailed investigation when an accident occurs and implementing the measures to prevent a recurrence based on the results of the investigation. However, in the current accident investigation, a more efficient and effective method was required from the viewpoint of the time required for the initial investigation, the preservation of the site situation, and the information sharing after the investigation. In this study, we applied SfM multi-view stereo photogrammetry, which is one of the image analysis techniques, to the actual train derailment accident investigation and examined its applicability. As a result, it is possible to obtain a detailed 3D mesh model of the site only from the photographs taken at the site, and it was possible to confirm the traces of the accident and the track members used in the site by utilizing the model. Furthermore, the created model has sufficient accuracy for measuring the position of traces of the accident and measuring the sleeper interval, confirming that this method is effective for improving the efficiency of accident investigation.

This paper proposes an AI image analysis system that can operate accurately in real time in an outdoor environment. The AI system receives the image from the network camera and determines the degree of danger according to the intruder detection algorithm. The judgment result becomes a trigger signal, and the rotating light of the alarm device is activated. The risk determination is controlled based on the state machine so that the alarm operation does not react excessively. Through a demonstration experiment at the Higashiyukigaya Works, it was confirmed that AI identifies people and vehicles in real time and that the function to prevent contact accidents works effectively.

Reinforced concrete (RC) structures in chloride-laden environment may have steel corrosion and corre-sponding corrosion cracks due to chloride attack. Although observed corrosion cracks provide an effective information on the status of steel corrosion inside concrete, their relationship is very complex because of many associated parameters such as structural details (e.g. cover and rebar arrangement). In this study, steel corrosion distributions in longitudinal and transverse directions have been obtained through corrosion ex-periments of RC members with different structural details. Parameters to represent the 2D-stochastic field associated with the steel corrosion distribution were identified. With the aid of 3D-finite element analysis, a database consisting of the relationship between steel corrosion and corrosion cracks for machine learning (pix2pix) was developed numerically. Finally, given corrosion crack distributions observed in the deterio-rated RC members and structural details, the structural capacity can be estimated.

Digital transformation (DX) of infrastructures field has been accelerated recently, and data linkage of maintenance data with BIM /CIM data is under investigation for their future use. On the other hand, the development of various sensing and monitoring technologies is in progress continuously, and it is likely that data that were not envisioned at the time of design could be effective for maintenance and management applications in the future. When implementing such new technologies in society, if the specifications for new data and API access methods are not consistent, data access methods will vary as the amount of data increases, and this is expected to hinder the sustainable development of DX. In this study, we focused on the convenience of schema-driven API development, such as the fact that data specifications and API access methods can be released simultaneously with data distribution and that new data can be easily provided as microservices without waiting for updates of existing systems, and examined ways to improve the efficiency of API. We studied how to improve the efficiency of API development for infrastructure data access.

Effectiveness of vibration measurement using acceleration sensors and anomaly detection with autoencoder was investigated, for improving non-destructive inspection and structural health monitoring for civil structures. In particular, a seismic damage detection method was examined in quasi-static loading tests on RC beam and column specimens. Local through-thickness vibration testing and acceleration monitoring with applying whitenoise signals were conducted on the specimens. Measured data of the intact specimens was used to train an autoencoder. Anomaly scores were evaluated for measured data of the damaged specimens, with the trained autoencoder. As a result, the shear crack, flexural cracks, and spalling of cover concrete in the RC beams and column were successfully detected. Furthermore, it is suggested that the proposed method with multipoint measurement has high possibility to identify locations and degrees of damage in RC members.

In Japan, we are aiming to realize a next society in which cyberspace and physical space are highly integrated toward the Society 5.0 era. There is a lack of human resources involved in bridge inspection or management and also Ministry of Land, Infrastructure, Transport and Tourism said to change the how to work called by i-Construction. In the maintenance of bridges and other social infrastructure structures, appropriate visual inspection results and evaluations are important for renovating and reinforcing bridges. In order to change these problems, association of public group or private company are held in seminar for bridge inspection for newcomer inspector. In the bridge-field site, staff and participant have been taken of car accident and weather, etc. Furthermore, staff and participant have to be mobile the field. For example, it takes for 1 to 2 hours by car depend on area of prefecture.

Recently, it is becoming to be popular for VR and AR techniques, which are collectively called XR. Our research group is made 3D-VR model by existing bridge and virtual bridge in order to learn how to inspection for newcomer bridge inspector. And we are also developing a system to display bridge information using AR in order to improve inspection work efficiency and to reduce hard print of documents. In the future as the 3D VR-model is used by MR-HMD.

In this study, it is aim to be efficiently at the bridge inspection work and management work. Firstly, we reported how to make 3D VR model, secondly how to make AR system and how to recognize existing model. According to these results, it is applied to learnt to how to inspect for bridge using 3D VR-model and AR system. Furthermore, in the near future these methods will be contribute to manage of civil infrastructure such as bridges.

Based on 675 natural water content data obtained from a peaty ground in the Ishikari River basin, Hokkaido, the probabilistic distribution of the natural water content has been modeled as functions of the depth using generalized linear model. Combination of three kinds of probability density functions for the water content distribution and six polynomial expressions for the mean from zero to five orders in terms of the depth have been examined for searching the best model. By comparing the 18 models from statistical and engineering viewpoints, the best models for organic and inorganic soils have been proposed, respectively. The validity of the models has also been confirmed. To estimate the in-situ natural water content distribution, the confidence intervals for the mean, 5 percentile, and 95 percentile points of the water content distribution of the proposed models have been calculated by employing bootstrapping method.

Pavement inspection is required to detect and report possibly collapsing spots as rapid as possible. We previously developed a GPR data analysis tool using AI to accelerate the data analysis process. We further developed a system that automatically transfers data and performs the analysis using AI on a cloud computing platform. The experiment with approximately 3 km of the survey showed that the data of sensors and results of the AI analysis became remotely available in approximately 36 minutes. The result suggested that the system is capable of significantly shortening the time required for remote inspectors to begin their tasks and reducing manual operations, which yields a significant improvement in the work efficiency of the overall pavement inspection project.

s automation of crop sorting has been widely implemented due to the decrease in the farming population, complete elimination of potato common scab tubers is required especially in the sorting of seed potatos. In this study, we aimed to detect the area of common scab on the surface of potato tubers by inputting two types of images: a conventional color image and a near-infrared (NIR) image at 960 nm. The common scab areas were manually labeled, and the segmentation model based on semantic segmentation was compared with a conventional model based on principal component analysis and support vector machines (PCA-SVM). The results showed that semantic segmentation showed higher accuracy than PCA-SVM, and the common scab areas were almost detected. In addition, higher segmentation accuracy was obtained with four inputs of RGB and NIR images than with only color images, suggesting the potential of NIR image input for common scab segmentation.

This study analyzes the relationship between vacant house occurrence and change of regional characteristics and living convenience. This analysis has been conducted focusing on the region in Kochi city where sprawled during the period of high economic growth in Japan. To understand the tendency, vacant house data set is prepared based on digital housing map which is commercially available and vacant house occurrence is analyzed based on prepared data. It was clarified that prepared vacant house data has similar characteristics with the survey results by local government. It means this study also clarified prepared vacant house dataset can be utilized to understand changes of vacant house occurrence. In addition to that, factors influencing vacant house occurrence statistically analyzed based on discriminant analysis. Explained variables were prepared such as location of convenience facilities, public transport services, population and specify of Densely Inhabited District (DID) in study site. Based on parameter estimation results by discriminant analysis, it is statistically clarified that variable which represents population and number of households are significantly affected for vacant house occurrence. In addition to that areas where specified as DID during high economic growth period and withdrawal of convenience facilities which is mainly used by specific customers tend to be higher ratio of vacant house occurrence.

In recent years, neural networks and deep learning have been studied in the field of coastal engineering. In particular, deep learning has been widely used in wave prediction because of its superiority in computational cost over the numerical simulations that have been used in the past. However, deep learning in previous studies has mostly focused on pinpoint prediction at a single point, and little research has been conducted on the prediction of areal wave fields. In this study, we propose a method to obtain the spatial distribution of waves by deep learning using a weather field forecasted by numerical simulation as input. The proposed method improves the accuracy by introducing a weighting to the loss function to eliminate the influence of the boundary conditions on the learning over land. In addition, the accuracy of wave prediction was improved by learning meteorological fields at multiple times as input data.

Mathematical models are widely used to represent the cyclic shear behavior of soil based on cyclic shear test data. These models, however, cannot accurately trace all test results. In this study, the researchers de- velop a new model that combines the advantages of both a deep learning model that can accurately repro- duce test data and mathematical models that robustly represent unknown behaviors. This combined model shows not only an improvement in the prediction performance for shear stiffness at increased levels of strains, but also more robustly represents unknown behaviors by virtue of following mathematical models.

Bayesian optimization has attracted much attention in recent years as a promising method for black-box optimization problem. In this study, we introduced Bayesian optimization to the centroid moment tensor inversion, which is one of the black box optimization problems in seismology. Through a synthetic test simulating real earthquakes, the analysis using Bayesian optimization was able to reach a solution closer to the true with fewer trials than that using random search.

In winter, a lot of heavy snow events occur on the north side of Japan, because of significant meteorological characteristics of a convergence zone, Japan sea Polar air mass Convergence Zone; JPCZ. The weather forecasts about the location and the strength of the JPCZ are difficult, because the development of the snow clouds are very rapid and the convergence zone fluctuates with instability. As inherent characteristics of the snowfall around the JPCZ, the snowfall contaminates the water as well and the snow over the highways results in car accidents and stalled cars. Therefore, the East Nippon Expressway Co. Ltd. (NEXCO East) seeks to optimize the operating strategies of deployment, recombination, and rotation about multiple groups of snow removal in multiple bases. This optimization requires some important indexes in combination with the weather observations and forecasts. For example, the reformation of the snow removal groups needs the information about heavy snow needed with an additional equipment and light or no snow without any equipment. NEXCO East and WNI with the University of Oklahoma are researching an algorithm and analysis system since 2019. The most challenging part of this research is to improve the snow prediction, because there is a mismatch between numerical simulations and real - time observations. Therefore, the forecasts both with observational infrastructure such as meteorological satellites, operational wide - range weather radars, and polarimetric radars, and with frequently - updated analysis and forecast system through the artificial intelligence, need to be developed. As a result, we are supposed to build the forecasts beyond a statistical analysis. In this paper, we'll discuss the initial results of observation - based forecast systems and show the relationship between real - time estimation of snow amount and classification aloft, and the optimization of snow removal.

Drain of rain water plays a very important role for the prevention of corrosion for either of steel and concrete structures. This study aims at developing water leakage detection system. In particular, this study focuses on the elbow of the drain pipes. Some experiments were carried out with the cracked and no-cracked specimens of poly-vinyl chloride water drain pipes. An infrared camera was employed to take thermal images. Four algrithms of the machine learning were applied. Those were random forest, AdaBoost, XGBoost, and convolutional neural network(CNN). AI learners were established by about 500 teacher data. AdaBoost showed the best performance of F value of 0.75, and CNN showed the second best performance of 0.73.

In Japan, water quality accidents caused by oils and chemical substances spilled from factories have occurred. Water quality monitoring is carried out by visual confirmation by facility managers and water quality inspectors at wastewater treatment facilities and sewage treatment plants , but continuous monitoring is necessary regardless of day and night, and there are problems with the allocation of monitoring personnel and the monitoring method of water quality. Therefore, in this study, we built a model that automatically detects water quality abnormalities using deep learning, which is a type of AI technology and has advanced image analysis ability. Feasibility Study conducted at wastewater treatment facility．As a result, we showed the possibility that deep learning can be an effective technology for improving the sophistication and labor saving of water quality monitoring.

For the purpose of confirming the applicability of the local outlier factor method (LOF) in non-destructive inspection to detect shear damage inside reinforced concrete beams，an experimental study of RC beam specimens was conducted. When shear loading was performed with a small load that did not cause visible shear cracks due to shear damage inside the RC beam specimen, the time history of the hitting sound and the concrete surface vibration acceleration at the time of hitting with a hammer was recorded. We tried to judge the damage by the judgment method based on LOF using the spectrum obtained by Fourier transforming the recorded time history. As a result, it was confirmed that the presence or absence of shear damage could be determined using the proposed method, maintaining a true negative rate and a true positive rate of approximately 90%. In addition the threshold setting range required for LOF was discussed..

If the defect of anticorrosion function is found, it is desirable eliminate the cause. When it is difficult to eliminate the cause, repair coating will be examined. Then, surface preparation grade by blasting stipulated by ISO Sa2 1/2 or higher is required as surface treatment before repair coating. Evaluation the surface preparation grade is performed by visual inspection in comparison with the representative photo example of ISO 8501-1. In this study, the authors try to develop the support system to determine the surface preparation grade from the image of blasting and discuss the difference between the learning model and the judgment accuracy due to the difference in the size of the teacher image data. The results suggest that system judged the unevenness and hueinthe imagebyvisualization ofbasisfor Judgmentusing Grad-CAM. It wasconfirmedthatsupport systemclassified surface preparation grade based on basis of judgement clearly.

For the purpose of immediate damage detection over a wide area after the disaster, we created 26,938 training data, by visual interpretation using the WorldView-3 satellite images acquired immediately after the Kumamoto earthquake. The building damage is classified into 3 grades, and the presence of blue sheet covered building is interpreted. Next, 37,121 patch images were cropped from the satellite image into 128 pixel squares, the patch acquired from 80% of the area on the north side of the satellite image is used as the training data, and the patch acquired from 20% of the south area is used as the test data. Next, we developed a program that automatically extracts the building shape and automatically classify the damage and the presence of blue sheets using U-Net, which is a semantic segmentation method that uses deep learning. As a result of evaluating the model using the training and test data, the IoU of the building shape is about 64%, the average F-mesure of the three damage categories is about 74%, and the F-measure for the blue sheet covering is about 89%. It was confirmed the high damage extraction performance that can be used to grasp the degree and location of damage.

In order to study how to utilize open data on traffic, we downloaded traffic data of England, where the data are available to the public. We focused on short-time traffic congestion prediction as one method of utilization, selected a relatively congested area in southern England and used PyCaret, a type of AutoML, to predict traffic congestion. Our model performed slightly better than models that assumed that the current situation would continue as is or that only the day of the week and time of day were used as input variables, indicating that machine learning can be used to improve traffic congestion prediction. On the other hand, the prediction performance of the as-is model varied greatly depending on the direction of travel at the same location, and the performance of the machine learning model also varied significantly accordingly. In order to compare the performance of machine learning traffic congestion predicts, it is necessary to establish a baseline forecast and show the improvement in performance against that baseline forecast.

The Ministry of Land, Infrastructure, Transport and Tourism (MLIT) aims to introduce ICT for construction sites and utilize three-dimensional data. The MLIT releases data acquired at the stage of planning, investigation, design, construction, and maintenance of civil infrastructures in MLIT Data Platform. However, it is difficult to use a uniform method for data acquisition, because the scale and conditions of construction sites vary widely. And, there are no established method for transferring three-dimensional data among the various project phases of civil infrastructures and for its specific utilization method. In this study, the concept of collecting, processing, managing, and utilizing construction and maintenance management information was proposed based on the definition of information system for constructing and using threedimensional data at small and medium construction site. We verified its application to progress management at the construction site.

The surface preparation grade of steel bridges is generally judged by visual inspection. There is a high possibility that a difference in evaluation by the engineer. The authors are trying to develop the support system for classify the surface preparation grade using deep learning. In order to improve the generalization performance of the system, it is necessary to enhance the teacher images used for training. In the present paper, it was confirmed that the generalization performance of the support system for classify the surface preparation grade can be improved by enhancing the teacher images with images generated by GAN. The SSIM value confirmed that mode collapse can be identified from the similarity between the teacher image and the generated image by GAN.

This paper presents a winter road surface condition classification method using deep learning with focal loss based on text and image information for detecting the deterioration of road surface conditions caused by snow accumulation. The proposed method achieves multimodal road surface condition classification by constructing a deep learning model that can cooperatively use images automatically captured by fixed-point cameras installed along the road surface, and text data related to road surface conditions. Since the distribution of training data is biased toward winter road surface conditions, there is a concern that the classification accuracy may be degraded due to the data imbalance problem. Therefore, the proposed method uses focal loss, which can deal with data imbalance, to train a deep learning model to realize road surface condition classification considering data imbalance. In the end of this paper, we demonstrate the effectiveness of the proposed method by conducting experiments using real data.

The inspections of tightening completion of high-strength bolts have been conventionally conducted by visually checking all the bolts. In this study, an inspection system using image recognition AI technology was developed. This system automatically identifies the fastening completion status based on still images captured by a smartphone. We conducted a demonstration experiment using this system and confirmed that the system can accurately identify the fastening completion status, and that it is effective in improving efficiency.

This study is intended to discuss the feasibility of the frequency identification and anomaly detection of highway pole structures from video footage. The accuracy of the identified frequencies from video footage is investigated through a laboratory experiment on a full-scale pole structure. Loosening anchor bolt conditions are considered as a damage scenario. In order to extract vibration modes of the pole structure, this study applies Phase-based motion magnification method which processes to the whole image frame and expands the designated frequency bands. Then, vibration waves are extracted by tracking the featured points in the image frame. Although the accuracy of estimated frequency identified from the vibration modes extracted from the video footage still needs improvement, the decreasing tendency of the frequency due to the damage scenario was observed. This study also investigates ways to remove background in the video image for the purpose of reducing influences of noise and improving identification accuracy.

The July 2018 heavy rainfall in the Western Japan caused extensive disasters over a wide area. From a long-term perspective, the occurrence of such heavy rainfall is on the increase, and it is important to predict where landslides will occur. In this study, we developed numerical models for predicting the locations of landslides using random forests, a machine learning technique. Two models with different explanatory variables were developed, and their prediction accuracies were compared. In addition, we also used SHAP, a type of explainable artificial intelligence (XAI) that has been studied in recent years, to provide global and local explanations of the numerical models.

In recent years, sediment disasters caused by heavy rains and typhoons have occurred frequently in Japan. Especially within the Fuji River Basin in central Japan, many regions are vulnerable to sediment disasters. For the risk reduction of sediment disasters, it is necessary to develop a sediment disaster risk estimation method that considers both trigger and inherent factors.

This study developed and validated a method to estimate sediment disaster risk that directly considers trigger and inherent factors in the areas around the Fuji River Basin (Yamanashi and Shizuoka prefectures) using a clustering and deep learning method. A fully connected deep neural network was used as the deep learning method. As the input data for the trigger factor, 60 minutes accumulated rainfall and soil water index within each cell were used. The horizontal resolution of cell size was around 1km. As the input data for inherent factors, the maximum slope angle and presence of faults within each cell were used. The sediment disasters caused by typhoons on 6th September 2007, 21st September 2011, and 12th October 2019 were used for training, determination of threshold, cross-validation, and validation of the deep learning method. For quantitative validation, we estimated the sediment disaster risk from neural network outputs and validated it by comparing sediment disaster occurrence reports. In the validation, an administrative area is regarded as "True Positive" if there are any" high-risk cells" and any "confirmed disaster cells" in the same area. We calculated the evaluation metrics based on a confusion matrix for validation.

Our method shows accuracy (0.314), indicating that the estimated sediment disaster risk is adequate. Therefore the risk information can help the decision-making of evacuation in the future. However, the False Alarm Rate was still high (0.904), so further improvements are required in future studies.

Diagnosis of Interstitial pneumonia by systemic sclerosis utilizes manual CT interpretation, in order to judge state of the disease and, thus, supports by computers is expected for that interpretation. In this paper, we develop an application to support the diagnosis of CT images. At first, we construct a neural network to accurately and robustly extract lung regions and disease regions from CT images by using deep learning techniques. Then, ratios of the disease regions toward the lung regions on all CT images of several patients are computed, and they are compared with results of other medical tests. Consequently, DLCO has a relatively close relation with those ratios. Finally, some useful functions for post process are also developed and they are integrated into an application for trial uses in medical practice.

This paper presents the development of a learning model to judge the Alkali Silica Reaction in concrete bridges from image data using the CNN algorithm. Gray scaling processing and brightness adjustment improve the accuracy of the judgment compared with color-image-based CNN learning. The learning model combined additional information other than image information, such as river system, name of structural components, and so on. As a result, the authors developed the learning model with 86.7% classification accuracy.

In order to improve the efficiency of infrastructure maintenance and management for port managers, the National Institute of Land and Infrastructure Management,to which auther belongs, is developing an inspection and diagnosis system for port and harbor facilities using UAVs and AI. The system uses UAVs and AI to convert port facilities into 3D data, and the target of the system is the automatic detection of facility deformations. This paper describes the AI-based detection of sea surface and sky and the AI-based detection of rust and exposed steel bars in the system.

It has been pointed out that a seismic isolation structure may collide with a retaining wall when a larger earthquake motion than assumed in the design is input. In order to simulate the behavior of the whole seismic isolation system when a seismic isolation structure collides with a retaining wall, it is necessary to model the restoring force characteristics of the retaining wall with the collision. However, for that purpose, nonlinear analysis by full scale experiment and precise three-dimensional FEM is required, and the modeling is not easy at all. Therefore, this paper proposes an evaluation method which can immediately calculate the restoring force characteristics of a retaining wall in a base isolation layer, which is necessary for analysis in the case of a retaining wall collision, from design parameters by using a machine learning prior learning model. In this paper, we first explain the proposed method, and then construct a pre-learning model according to the proposed method. Finally, this paper compares the restoring force characteristics obtained by the prior learning model and the restoring force characteristics made using the result of the three-dimensional FEM model, and shows the applicability of this proposed method.

Computation cost of particle filter, which is a representative method for data assimilation method, is generally large. Some problems in real world require real-time prediction by data assimilation. In this paper, we investigate a method to reduce the computation cost of particle filter for the prediction of curling stone trajectories for rapid data assimilation before and during a game. In order to reduce the number of parameters for data assimilation, trajectory angle and y-coordinate are simply calculated by a least square method. In advance 100,000 trajectory scenarios are calculated and scenarios which stop around the location where the stone stop actually in measured data. The likelihood is calculated only for the extracted scenarios. By these procedures, the computation time could be successfully reduced to less than 1 second.

The width of corrosive cracks caused by chloride induced deterioration in RC structures can be easily and massively measured by image processing technologies recently. If the spatial distribution of the corrosion of the internal reinforcing bars can be accurately estimated by these information, it would be useful not only for evaluating the current status of the structure but also for predicting the future of deterioration. In this study, we propose a methodology to estimate the spatial distribution of mass loss ratio of the rebar using Gaussian process regression, which is a probabilistic regression method, based on the corrosive crack width. In the estimation by Gaussian process regression, the model of cross-correlation is important, and the parameters of the model are estimated by the maximum likelihood method. In order to verify the proposed method, the estimated spatial distribution of mass loss ratio of the rebar is compared with the true one obatained by laboratory experiments. It is shown that the estimated distribution is generally consistent with the true one.

To improve the efficiency of bridge inspection, the automatic damage detection from images is expected. In previous study, we used close-up images of damage for learning data of automatic detection, but in actual bridge inspection, the damage detection is carried out for the image in which presence and position of the damage are unknown. In this study, we constructed an automatic damage detection model using deep learning and detected multiple damages on the whole bridge pier image. The detection result of the damage were displayed, layer by layer, on the bridge image. And we interviewed the bridge engineer to determine the damage. As the result, the positional relationship of the damages could be grasped at a glance by displaying the damage detection result on a whole bridge pier image. And switching of layers and on / off of the display of layers were effective for the damage determination.

Japan has taken the lead to realize Society 5.0 ahead of the rest of the world. This aim at innovations of the fourth industrial revolution (e.g. IoT, big data, artificial intelligence, robot, and the sharing economy) into every industry and social life. Especially in con-struction fields, improving productivity and safety management are expected using IoT and AI. Therefore, we have been focusing on the safety management of construction sites, and then we have proposed a personal identification method by deep learning using patterns on workers' helmets. In the present paper, to apply the past method to actual sites, we devised and evaluated a new correction approach with multiple cameras. Finally, we confirmed that the new method was useful to actual sites.

In this study, the image analytical method to extract aggregate regions and judge the joint disposal extent is established for the joint disposal judgement by the green cut for the joint of concrete such as concrete dam. In this method, aggregate regions are extracted and the disposal extent is judged by using area information judged as aggregates by using two kinds of smoothing images and the difference image between these images. Appropriate aggregate extractions and judgements were carried out by the examination by using taken images for joints in the experimental field and the site by using developed image analysis method. Furethermore, good agreements were obtained in the investigation by using panorama images.

3D modeling technology using point clouds needs further development to improve the efficiency of operation and maintenance for civil engineering structures. In this study, we propose a deep learning model and a subsampling method to estimate the longitudinal direction of bridges, which is necessary information to align point clouds in 3D modeling. Our deep learning model was developed by applying the pose estimation approach of the past study. Our subsampling method can generate datasets for deep learning as training data considering disturbance of point clouds with actual site conditions. The results estimated using our method were compared with that estimated using the principal component analysis, which is one of the conventional methods. Our method yielded higher accuracy than the principal component analysis even when handling point clouds with a large amount of noise and missing points. This means that our method has robustness for estimation of the longitudinal direction of bridges.

Polymer impregnated concrete (PIC) has excellent durability against impact and abrasion and is applied where robustness is required. An important indicator for measuring PIC durability is the impregnation rate of the polymer. A conventional method for measuring the impregnation rate involves comparing the weight before and after impregnation. While this method of comparing weights is straightforward, it requires extensive weighing equipment to be applied to large PIC forms. Such weighing is a challenge in terms of time and risk involved because it requires tasks such as lifting. In this study, we hypothesize that there is a specific relationship between the impregnation rate and the sound propagation characteristics. A device consisting of a speaker and a microphone was prototyped, and variations in the sound propagation characteristics were estimated using machine learning. The results show that the impregnation rate can be estimated with practical accuracy.

This study developed a steel corrosion monitoring system using a magnetostrictive vibration power generation device that converts energy from bridge vibration or wind vibration into electric power for deteriorated bridges due to chloride attack. In addition, as an application of this monitoring system, an analytical model was constructed by linking the monitoring results with a three-dimensional finite element analysis. And the residual load capacity was evaluated. Furthermore, this study linked the existing database system with real-time viewing of monitoring data.