In order to solve the shortage of laborers in the construction industry, it is expected to develop technologies that contribute to productivity improvement and emergency recovery systems that can cope with frequent natural disasters. With this background, the authors are participating in the Moonshot-type research and development program of the Cabinet Office, Japan, and have set the goals of emergency recovery from natural disasters and construction of lunar infrastructure using robotic technology, and are promoting research and development of robotic technology that can adapt to various environments and construct infrastructure. We believe that the key concept for such research is the "open design" of robots. In this paper, we introduce the outline of this project, and introduce the concept of "open design" and the research and development being conducted based on it.

The introduction of AI is being considered in various fields, and it is expected that AI learned from big data will be put to practical use. However, it is thought that various problems have arisen in the utilization of knowledge, not limited to AI, by applying disparate information on the Internet according to the user's situation. It is said that there is a need for structuring. Even in the utilization of AI in infrastructure mainte-nance, it may not be easy to use AI learned from data in practice, so it is necessary to structure the knowledge of infrastructure maintenance. In this paper, the structured knowledge in infrastructure mainte-nance is called specialized knowledge. From the standpoint that it is difficult to consider how to utilize AI without clarifying the expertise in infrastructure maintenance, refer to the analysis of the history of mainte-nance of the Japanese National Railways and practical knowledge in other fields such as nursing. After clarifying the expertise of maintenance, we examined how to utilize AI by linking deductive methods such as rule base and inductive methods.

Vision-based techniques for structural inspection and monitoring have seen significant advances, stimulating innovations in the area. Field implementations of such new methodologies require preliminary investigations, in which the applications of the methods are simulated. Those investigations facilitate the user’s understanding of the methods, and enable the quantitative evaluation and optimization of the method performance. However, specifications and requirements for vision-based inspection and monitoring methods differ significantly, depending on the target structures, environmental conditions, and the expected output data. Analysis and evaluation of new methods under individual application scenarios are therefore not straightforward. In this paper, the authors focus on the computer graphics simulations of structures as a potential solution to the problem, and discuss the application of such simulation to vision-based inspection and monitoring scenarios with examples, as well as the roadmap for the future extensions.

Inspection methods using new technologies such as deep learning have recently attracted attention as an alternative method of previous close visual inspection of bridges. New technologies which automatically detect damage from the image recognition of the bridge support bridge inspectors to diagnosie the damage, and the inspection work will be lobersaving. However, it is necessary to have the detection display accuracy which can make the decision equivalent to close visual inspection. In this study, free lime which is one of inspection items of the bridge inspection was automatically detected by the image recognition, and we verified the accuracy of the display result necessary for the damage decision of the engineer. In addition, we propose an appropriate display method for automatic detection of free lime as diagnostic information for bridge diagnosticians.

Convolutional Neural Networks are biased towards recognising textures. Thus, the accuracy is also reduced between images with different textures in crack Segmentation. In this study, We evaluated the difference of texture by calculating features of crack images. To make a model that can detect cracks more stably, we compared the performance of a model with only batch normalization and models contained instance normalization.

When steel corrosion occurs in reinforced concrete (RC) members, corrosion cracks appear on the con-crete surface. Although corrosion cracks are valuable information for predicting the amount of steel corro-sion, corrosion crack width depends on the corrosion amount of not only the closest rebar but also the surrounding rebars. This results in the complicated non-linear relationship between corrosion crack and steel corrosion. In this study, using an artificial database and pix2pix, 2D steel corrosion distribution is estimated given the distribution of corrosion crack width. As an illustrative example, considering model uncertainty involved in the prediction of steel corrosion, probabilistic density function associated with flex-ural capacity of aging RC beams is estimated conditioned upon the corrosion crack distribution.

Reinforced concrete (RC) structures should be designed with considering the effect of the surface chloride ion concentration (C₀). RC structures for port and harbor facilities are especially affected by C₀, and the value of C₀ for their design is estimated by the equation which is the linear regression of measured C₀ data with only the distance from the high water level in Japan. However, the equation is not sufficiently reproducible for all measurement data, and thus this study aimed to develop a new regression model by using machine learning technique, which is a powerful data driven approach. From the environmental factors, the variables contribute to the improvement of regression performance were extracted and the model was constructed by learning these data. The performance of the model was investigated and the result showed that “significant wave height” is useful for predicting C₀, for especially near the sea level.

The deterioration of infrastructure is a social issue. Street lights are no exception, and accidents are oc-curring in various places. The main in-spection currently performed is visually inspecting. When abnor-malities such as corrosion are found as a result of visual inspection, ultrasonic testing, magnetic powder testing, or penetrant testing is performed. However, none of these methods obtain continuous data, and it is difficult to detect cracks early and to predict growth of it. Therefore, mechanization and automation are required. To solve this problem, a con-stant street light monitoring system is proposed in this study. Re-garding the method of crack detection, the data is measured by using a measuring apparatus using audible sound for the Street lights. The power spectrum is calculated using the data, and deep learning, which is one of machine learning, is used to detect cracks early and to predict growth of it. In the experi-ment using the test body, the accuracy of the actual crack length and the predicted value of machine learn-ing was 63.6%, and the presence or absence of the crack could be detected with a certain degree of accu-racy. This result indicates that is able to considered practical enough.

When a train repeatedly runs on a track, the track irregularity, which is the distortion of the track, grad-ually increases due to load. Normally, the track irregularity is inspected periodically, and maintenance is performed when large track irregularity is detected. However, in rare cases, the track irregularity may pro-gress locally and rapidly. In order to ensure the safety of train operation, preventive maintenance is required to detect the signs of such rapid progress and to perform maintenance before it occurs. In this study, we have developed a model to identify in advance where large track irregularities are likely to occur by apply-ing cluster analysis to historical data of track irregularity and maintenance records. In addition, by applying this model to each track inspection and comparing it with the result of the previous track inspections, we showed the possibility of detecting sections of the track that require attention in advance. Specifically, the model extracts the section that have approached the center of the cluster requiring attention and the section that have moved between clusters.

Even if high-quality materials are used and properly mixed, the full potential performance of the concrete will not be exhibited if the concrete is improperly placed and compacted. Conventionally, the degree and the completion time of concrete compaction have been determined by visual judgment and feeling based on the experience of engineers. In recent years, with the decrease in the number of engineers and work style reforms, it is required to improve the productivity of concrete construction, and it is desired to develop labor-saving or unmanned technology while ensuring the quality of concrete. The authors have devised and verified a system in which AI can replace the conventional concrete compaction judgement of engineers. As a result, it is shown that AI can realize the concrete compaction judgment close to that of the engineers by learning the training dataset consists of pairs of the frame image taken by the video camera and compaction judgment of engineers, and its usefulness was comfirmed through evaluation experiments.

The deteriorated infrastructures constructed during the high economic growth period result in the need for enormous maintenance and management costs. Maintenance of the headrace tunnel using AI technology is desired in view of the points that the difficulty to deploy the large machine and water is cut off during inspection. In this study, chalking points on the inner wall of the headrace tunnel were detected by YOLOv5 in consideration of these situations. As a result, the chalking points can be detected with high accuracy by using the multiplier or clear data. In addition, it will be possible to investigate the causes of deterioration and propose the repair plans by understanding the distribution of deteriorated areas based on the detection results since a part of the chalking area can be captured even the IoU is low or the detection is not accurate. For the classes that were not detected, it is necessary to consider the augmentation of training data for infrequently occurring deterioration or use the classification model added to the model used in this study in order to improve classification accuracy.

AI research and application in civil engineering are being promoted for improving the efficiency and labor saving of various operations. While some technologies such as applications that determine damage level and its extract area from images have been developed in the maintenance field, there are some fields where AI research has not progressed very much. In this paper, we describe the current status, issues, prospects, and expectations of AI application for steel structures. First, the current status of AI research and application for steel structures are compared with that for the other fields in civil engineering. Then, the issues to be solved are pointed out. In the section on prospects and expectations for steel structures, by referring to the AI research and application in various fields, we presernt the future direction of AI application and ideas to activate the study in the field of steel structures.

As the damage of bridges is expected to accelerate, there is a need for maintenance and management based on an efficient inspection. The Convolutional Neural Network (CNN) is a type of the machine learning, which is one of the most effective methods to support the inspection of steel girder bridges. However, the computational cost of a pixel-by-pixel corrosion detection such as the semantic segmentation method is high, and it is difficult for inspection engineers to use that method in the inspection site. Therefore, the purpose of this study is to propose our method that allows inspection engineers to detect a corrosion and its area at short times in the inspection site. A series of methods in this study were which examined, we first performed the corrosion detection using the CNN, and then the corrosion area calculation applied the image binaization to the output of the CNN.

As city vitalization through the utilization of public space is getting attention, the need for observation of human behavior in urban space is increasing. However, the analysis work relies on manual video observation, which requires an enormous amount of labor. In this study, we applied object detection and tracking technologies based on deep learning to develop a low-cost human flow analysis system, especially for outdoor urban spaces, which enables easy and fast analysis. Using the existing video data and human flow analysis data from the experiment in the Mitaka station area, we verified the analysis accuracy of the proposed method by setting "mesh" and "rounding threshold" for practical use in the developed system. As a result, we confirmed that the proposed method can grasp the trend of human flow in 1/4 of the analysis time of human survey. We also clarified issues for practical use, such as the method of camera installation and the necessity of image distortion correction.

As of 2021, the spread of COVID-19 infections continues worldwide.COVID-19 is not only life-threatening, but also affects the politics and economy of many countries.In addition, it has brought about changes in people's daily life behaviors, such as commuting to work or school, receiving medical treatment, and purchasing.In this study, focusing on changes in medical treatment behavior and, using National Health Insurance Data to examine changes in outpatient medical treatment before and after the spread of COVID19 infections.National Health Insurance Data has been accumulated on a monthly basis since April 2012, and receipt information, including medical cost information, medical treatment information, and disease information, has been recorded for each individual.Through the analysis in this study, it was found that there was a clear change in the outpatient medical treatment behavior of local residents before and after the spread of the COVID-19 infection.

We have optimized the parameters of dam inflow prediction models based on certain observation data and have thrown them into real sites until now. On the other hand, we have generally thought the input data into models as uncertain predictive rain given by Japan Meteorological Agency. There are differences in data character between observation and predictive data, so we cannot expect the improvement of dam inflow prediction precision until the advance of predictive rain precision. Therefore, we offer to construct and learn models including predictive data to minimize data differences between building and operating models. In this thesis, we call this process Prediction Learning. We indicate that Prediction Learning under uncertain circumstances of operating conditions enables dam inflow prediction in operation to predict in high precision. And we show that civil engineering interpretation adding to predictive data allows the AI model to be easy, and dam inflow prediction precision one to six hours ahead is not decreasing.

Data science methodologies, which have undergone significant development in recent years, provide flexible representational performance and fast computational means to address the challenges faced by traditional scientific methodologies while simultaneously revealing unprecedented challenges such as the interpretability of computations and the demand for extrapolative predictions on the number of data. Methods that integrate traditional physical and data science methodologies are new methods of mathematical analysis that make complementary use of both methodologies and are being studied in a variety of scientific fields. In this paper, I point out the significance and importance of such integrated methods from the viewpoint of scientific theory, and survey and systematize specific methods and applications, and summarize the current knowledge in the relevant research fields.

Surrogate models, which reproduce the inputs and outputs of physical phenomena in a data-driven manner, are increasingly being used as an alternative means of making fast predictions of physical problems, but there is no guarantee that the solutions will satisfy the physical conditions. Physics-Informed Neural Networks (PINNs), on the other hand, are neural networks that solve the governing equations in a data-driven manner by introducing a loss function that represents the constraints imposed by the governing equations. In this paper, we present the formulation and code implementation of PINNs for a one-dimensional continuum free vibration problem.

Our study presents surface water temperature predictions on a reservoir using multiple climate model datasets as the inputs to recurrent neural network (RNN). We focused the reservoir in Hokkaido that is located in the mid-latitude region and is supposed to be strongly affected by future climate change. With past and future dynamical-downscaling data of the climate change data (Scenario RCP8.5 for future) to the reservior-located area, the RNN that was trained by observed data (temperature and water temperature) was run to predict the water temperature on the reservoir. The difference between past and future climate change data (i.e., Future minus Past) in each GCM showed that the future temperature increased by 2 to 4 °C in a monthly average, which supports RCP8.5 projection. This trend of the air-temperature difference was sim-ilar to the predicted difference by RNN in water temperature (i.e., the water-temperature difference between monthly averages of past and future was 0 to 1 °C). However, during spring and summer periods, the water temperature predicted by the past temperature was marginally higher than that by the future temperature.

In this study, we improved and verified the dam operation model using deep reinforcement learning. Previous studies of this model have only verified for the virtual floods, which greatly exceed the design floods, and have not verified their effectiveness during actual floods. In this study, we added condition settings for suppression of over-discharge and improved the reward function based on the problems of the previous model. In addition, by adjusting the flood scale of the learning data used for deep reinforcement learning, we aimed to acquire effective discharge operations for floods of the design scale. We compared the improved dam operation model with the the previous model for recent actual flood case which caused disasters. It was confirmed that the improved model reduced the peak discharge compared to the previous model and was similar to the discharge pattern of the actual dam operation..

Conventionally, physical models such as the tank model have been used as a dam inflow prediction method. Since a physical model can be thought as an approximate function of the actual phenomenon, it should be possible to predict performance equal to or better than a physical model if it is replaced by a neural network. Therefore, in this study, dam inflow predictions of a tank model and a neural network were compared under the conditions that input data were equal.The tank model of this study was able to predict the inflow amount at the time of large-scale flooding with relatively high accuracy by adjusting parameters using the latest observed values. On the other hand, the neural network trained with the same input data showed prediction accuracy equal to or better than that of the tank model. This result suggests that the lower limit of the predictive performance of a neural network is given by physical models.

This study presents the rainfall data analyses for cut-slope failure caused by heavy rainfall, based on the actual failure cases of the expressway in the Tohoku district. The characteristics of the rainfall data indicated that the Radar/Raingauge-Analyzed Precipitation by JMA has conformance for deep learning and that it is difficult to judge the failure probability of a cut slope only from the rainfall data as the triggers. In addition, the authors developed a deep learning model by combining incitements and predisposing factors based on the rainfall analysis to judge the failure probability of cut-slope. As a result, the deep learning model with 92-96% of accuracy and 65-74% of sensitivities was constructed. However, the developed deep learning model showed a tendency to judge cut slopes with high risk based on predisposing factors slightly biased toward the failure side, and that the necessity of accumulating more training data was clarified for the deep learning.

This paper presents a study on validation of the training data for the predisposing factors of the slope failure in order to apply deep learning to the estimation of the probability of slope failure of cut-slopes on expressways due to heavy rainfall. In the previous study, a method for quantitative evaluation of slope failure risk was proposed by artificially weighting the features of slope predisposition. By contrast, to improve the objectivity and versatility of the dataset, the authors constructed the supervised learning with the objective dataset of only items without weighting of features, and compared the model with those of previous studies in terms of validation. The results indicated that the accuracy was improved in the latest failure cases when the training dataset was not weighted.

It is important to predict the amount of precipitation that makes roads closed due to landslide disasters to reduce human damage and secure evacuation routes during heavy rain. This study attempted to estimate a prediction model for road closed that trained the data of heavy rain in July 2020 by using the passable route map as labels in machine learning. However, the estimated model could not predict the labels of road closed and it was found that the estimated model was not available for the prediction model. The reason for the result may be there were mistakes in the labeling process and that only the topographical information and rainfall information were not sufficient as input data.

For the purpose of damage detection immediately after a disaster, we developed a deep learning model using aerial photographs taken from an oblique direction with a helicopter or drone. This model automatically extracts damages to buildings and landslides, then divides into four classes: no damage, damage, collapse and landslide. As a result of discrimination using unlearned test aerial photographs using this model, it was confirmed that the average Fmeasure of each class was about 64% and mAP was about 0.35.

In Japan, there have been many landslide disasters caused by earthquakes and heavy rains. The Geospatial Information Authority of Japan (GSI) prepares maps of collapsed areas to assess the damage, but it takes a lot of time for technical experts to decipher the collapsed areas manually. In recent years, there has been research on the detection of landslides using machine learning with deep learning. However, the research is still in its early stages, and the methods and analysis data still need to be accumulated. In this study, we propose a method for detecting landslides using Semantic Segmentation based on deep learning, and compare the results of detection with different training data in order to improve the efficiency of locating landslides. In addition, this study aims to detect landslides using only post-disaster aerial photographs to efficiently understand the damage situation, which is also highly novel.

In Japan, large-scale earthquake disaster occurs frequently. When a large-scale earthquake disaster occurs, it becomes difficult to collect the victim’s needs and location. In addition, mobile phone and disaster prevention application cannot be used during large-scale earthquake disaster. The authors are developing a proposal to enable vulnerable people to signal their location and needs in the aftermath of a disaster to response teams by deploying radar reflectors that can be detected in synthetic aperture radar (SAR) satellite imagery. In this study, Information retrieval system for ascertaining vulnerable people to distribute the reflector is proposed. Through the analysis of this study, the Information retrieval system considering various physical condition was developed. In addition, regarding the system, useful evaluations were obtained from experts in disaster prevention engineering and health science.

In this research, a non-destructive inspection method for a debonding part of CFRP-concrete structures is investigated by using the laser ultrasonic visualizaiton testing (LUVT) and a deep learning. In general, time-stepping images (or movies) of ultrasonic wave propagation in the laser radiating surface can be obtained by using LUVT. At that time, LUVT inspectors have to give an evaluation on with and without defect by visual judgement from such LUVT time-stepping images. However, considering the common concern about the shortage of non-destructive inspectors in near future and the complex ultrasonic wave elds from the anisotropic property of CFRPs, AI alternation for this visual judgement of inspectors might decrease loads of themselves. Therefore, in this research, some CFRP-concrete structure specimens are prepared, LUVT inspections for them are carried out, and some images for them are obtained. Then, a deep learning is implemented in order to construct an AI for a debonding part of CFRP-concrete structures. As learning and testing results, it is concluded that a debonding part can automatically be detected by LUVT if the CFRP thinckness is thin.

Many mortar-sprayed slopes were constructed during the period of High economic growth, and they are aging at the same time. It is difficult to visually check for floats among the abnormalities of mortar-sprayed slopes, so percussion inspections are used to check for them. However, due to the shortage of inspection engineers and the financial difficulties of the national and local governments, there is a limit to the amount of time that can be spent on continuous diagnosis of floats using only percussion inspections. In this study, we developed a deep learning model using infrared images acquired from an infrared camera mounted on a UAV, taking advantage of the difference in heat capacity between a floating part and a sound part. As a result of experiments using images that were not used for learning, it was confirmed that the model could accurately estimate the floating areas on the mortar-sprayed slope.

It is important to enhance the migratory for the revitalization of the central city area. In this study, a variable including movement resistance was created for the central city area of Nagano City. In addition, it created policy variables that change depending on the revitalization plan. A pedestrian’s rambling activity model constructed using these two variables. In addition, it simulated the amount of pedestrians to estimate the pedestrian flow. From these results, it was clarified that the change of the pedestrian amount caused by the pedestrian’s rambling activity in the city area of Nagano City.

In recent years, It becomes the problem that a population decline, low birthrate and aging is social. In addition, it is necessary to solve the extension of the town by the development of the motorization. There-fore the Ministry of Land, Infrastructure, Transport and Tourism raises it by letting a central city area acti-vate when the realization of the collection type city is important. In this study, GPS data obtained from the smartphone application "TriPre" was utilized using the central city area of Kanazawa City as the research area. Then, the data change by the action extraction was analyzed. As analytical methods, movement locus and karnel density estimation were used. From the results of the study, changes in the sphere of action in transportation and day and night and COVID-19 were confirmed.

In many local cities, the government is facing a decline in the liveliness of the central city due to the declining birthrate, aging population, and motorization spiral. In order to revitalize the local area, it is im-portant to focus on the flow of people and to consider measures for citizens and tourists, taking into account the population staying in the area. In this study, we analyzed the changes in behavioral characteristics caused by the spread of the new coronavirus infection in four regions based on a questionnaire survey and mobile spatial statistics data with attribute information. The results showed that the behavioral characteris-tics of the new coronaviruses changed in four regions. The results showed that the regional characteristics of the four meshes differed depending on the time of day and period due to the expansion of COVID-19, and that the differences between the central city axis area and the residential area were particularly large.

The spread of COVID-19 infection has required a shift to a new lifestyle in Japan, and purchasing be-havior has changed as a result of the spread of the infection. In this study, we investigated the effects of the new coronavirus infection on purchasing behavior in Japan. In this study, we focused on the changing lifestyle, especially the purchasing behavior, and analyzed the municipalities in Ishikawa Prefecture using the big data of purchasing. The results showed that the purchasing area changed depending on the size of the municipality, and that the amount of money spent on purchases and the main items purchased changed before and after the spread of COVID-19 infection by focusing on the region, age, local purchases, and out-of- town purchases.

In this paper, a new structural optimization method by constructing an agent to play a game by reinforce-ment learning is proposed. First, the outline of the competitive structure optimization game is explained, and then, the composition of the agent and learning plan is described. In addition, by confirming the play result of the game by the constructed reinforcement agent, the possibility of structural optimization by the proposed method is verified. As a result, it was confirmed that there was a possibility to construct a structure optimization agent by reinforcement learning, though there was a problem in generalization performance.

When an earthquake occurs, the larger the scale of the disaster, the harder it is to support the victims’ needs. In fact, even the most meager support for victims has become quite difficult. Furthermore, it is also known that even with the use of mobile phones and disaster-prevention applications, the greater the earthquake disaster. The authors have developed multiple reflectors observable by synthetic-aperture radar (SAR) satellites and differing backscattering coefficients, and proposed having them set up by victims immediately after a large-scale earthquake disaster occurs, as well as a means to ascertain their location and needs. In this study, we propose a method for detecting reflectors observed in a wide area SAR image by combining the object detection algorithm YOLOv5 and an anomaly detection method using time-series SAR images. Through the analysis in this study, we were able to detect 8 reflectors out of 9 observed in a 10×10 km SAR image.

In recent years, many people who need support for evacuation have been affected by large-scale disasters that occur frequently in Japan. In this study, we conducted a survey of the evacuation behavior of people in Ishikawa Prefecture, Japan. In this study, we simulated the evacuation of people who need support for evacuation using an automated vehicle, assuming that the Ladder River overflows in Komatsu City, Ishikawa Prefecture, and calculated the time required for evacuation. The time required for evacuation was calculated. By changing the parameters of the time required, we were able to obtain knowledge about the time that should be reduced when the time required for evacuation support using an automated vehicle is shortened.

There are numerous in-service structures in modern society, and establishing an efficient management scheme is an urgent issue for the sustainability of our life. The recent digital transformation technologies such as SfM and laser scanning are promising for obtaining detailed as-built information of structures. We investigated the 3D digital modeling techniques for the maintenance of infrastructures especially point cloud data. A literature review of practically used systems for infrastructural resilience management in Japan is conducted. Then, we discussed current challenges and prospects for introducing a 3D model in the maintenance engineering of structures. From the above studies, both laser scanning and structure from motion (SfM) techniques are commonly used to construct point cloud data. UAV and other mobile robot systems enhance the applicability for large-scale infrastructures. On the other hand, the cost of the system and handling large amounts of data are challenging issues, and transformation from the 3D model to structural analysis is also another issue to utilize such information in maintenance engineering.

In this study, the prediction and regression analysis of shrinkage during 6 months of drying period were carried out for the database of drying shrinkage experiments of concrete by using the JSCE prediction model and several machine learning models. In the machine learning, the default model parameter settings and the models with optimized parameters were compared. As a result, the parameter optimization slightly improved the prediction accuracy but the database and the training model used in this study were able to predict at a certain level of accuracy even with the default settings. In addition, accuracy of the machine learning prediction of shrinkage for 6 months of drying significantly was improved when the shrinkage strain for 28 days drying period was used as a predictor. In the importance analysis of the predictors, the importances of unit water content and aggregate density on the prediction of concrete shrinkage with 6 months drying were high. When 28-day dry shrinkage strain was included as a predictor, the importance was the highest while the importances of aggregate properties were relatively reduced. It was suggested that 28-day dry shrinkage strain partially includes the influence of materials properties and other parame-ters on 6 months drying shrinkage, which would be critical for drying shrinkage increase even not taken into accout in the design parameter.

It is expected that the demand for big supply of various data of infrastructure, not limited in AI application field, will increase, and it is required to develop the efficient data supply scheme in order to provide more detailed data to data platform for future data utilization. In this research, we focused on the utilization of MQTT protocol as a method that can realize automation of measurement data supply and immediate multiple use, and improved the efficiency of automatic measurement data supply when combined with automatic data measurement by LPWA. To ensure security during data distribution, we used MQTT with TLS connection in addition to LoRaWAN with AES encryption method. We constructed a prototype for measurement and data supply, and examined the possibility of real-time supply of infrastructure data by taking advantage of the characteristics of MQTT, which is a widely used protocol for one-to-many connection.

COVID-19 was first identified near Wuhan, Hubei Province, China, in November 2019. It has since spread around the world, causing a pandemic. In order to prevent droplet infection through person-to-person contact and the influx of the virus into the country, countries around the world have implemented measures such as curfews and border closures. In this study, we attempted to understand the impact of COVID-19 on the aviation industry in detail. The data obtained from Flightradar24, an application that displays the current position of commercial aircraft in flight in real time, was used for the analysis. Through the analysis in this study, it was possible to obtain a time series of the number of aircraft operations due to the impact of COVID-19. In addition, the differences in the response to COVID-19 among airlines on the same route were clarified and discussed in terms of the number of passengers.

In the field of maintenance and management of social infrastructure, there is a shortage of engineers due to the declining birthrate and aging population. In addition, as the number of aging structures is increasing, it is necessary to reduce maintenance costs. In recent years, the ICT field, which is an alternative method for close-up visual inspection, has been attracting attention as a technology that can be expected to solve such labor shortages and reduce costs. With the method using VR (Virtual Reality), it is not necessary to go to the site, so labor saving in the inspection work of social infrastructure can be expected.line. In this research, we utilize VR technology to reproduce an actual bridge in VR space and display the damage detected using AI. Consequently, establish an inspection environment and verify the possibility of inspection support by displaying AI results

In this study, taking the case of local city in Ishikawa prefecture as an example, using the network analysis, the facility coverage rate of each town was scored based on the existence and type of facilities that satisfy the needs of living within a 10-minute walk from the facility, and the k-means method was used. We estimated the future of life vulnerability considering the change of population and the possibil-ity of withdrawal of facilities.

Comparing the life vulnerability between 2020 and 2045 as an analysis result, It is predicted that the fragile area will expand due to the shrinking walking range and the withdrawal of facilities due to the ag-ing and declining population. In addition, areas with high vulnerability to living will rise significantly. The result was that the area of sufficient life maintained the status quo, but this is due to the influx and concentration of vulnerable town characters into the center of the city, which is expected to reduce the size of the city itself.

The inspection and diagnosis of weathering steel bridges is based on the deterioration of anti-corrosion function, which is evaluated with the corrosion condition rating by observation. However, in order to perform an accurate evaluation by therating, inspectors need to have necessary to have adequate experience. Due to the recent shortage of human resources and inspection costs, it is needed to establish a simple and accurate evaluation method. In this study, the identification of rust condition has been investigated by using digital images of surface rust condition in weathering steel bridges and existing CNN models. Inaddition, the effect of digital image resolution on the identification accuracy has been considered. As a result, the accuracy isrelatively high in the case of VGG19 and SEnet. The larger input image size gives the better accuracy. Furthermore, the accuracy tends to decrease when the resolution of the images used for training and validation is different.

It is important to detect and locate damages in structures at an early stage to extend service life of the structures. Vibration-based damage identification methods that examine changes in vibration characteristics between intact and damaged stats of structures has been proposed by many researchers. However, it is difficult to accurately estimate multiple damages caused by natural disasters or long-term deterioration, for example. This study attempts to identify multiple damages in a cantilever and a simple beam using neural networks based on natural frequencies and natural vibration modes, which are typical vibration characteristics. To examine the accuracy of identification, particular attention is given to high-order vibration characteristics and slope elements in natural vibration modes. The results suggest that suitable data preprocessing method for each boundary condition is required to improve the accuracy of identification.

This study aimed to propose a method for the extrapolation of dam inflow predictions applicable to water utilization dams. In recent years, in response to the frequent floods that have occurred nationwide, improing the flood control function of dams through leveraging pre-emptive discharge has attracted attention. How-ever, as there is little observed information regarding water utilization dams and because specific flood control operations have not been decided upon, there is concern regarding the adverse effects of such ac-tions on water use. Therefore, difficult judgments on the implementation of pre-emptive discharge are re-quired, which increases the burden on managers. In tandem with this, there is a desire to improve the accu-racy of inflow prediction in water utilization dams. In this study, we propose a regression model that can predict the inflow of dams based on analysis using Elastic Net, a sparse modeling method that is utilized to identify relationships between data from small amounts of information. In summary, we were able to pro-pose a method to make predictions, even in circumstances with scarce information gleaned from observa-tion.

Estimating the amount of accumulated snow in mountains is important in terms of both water use and disaster prevention. However, limited number of observation point makes it difficult to graspe the snow depth distribution accurately and quickly. To address this limitation, we propose a deep learning method that estimates snow depth distribution from information which can be obtained in real-time, such as snow cover extent derived from satellite data and snow depth at observation points. Physical simulations that modeled the snow accumulation / snow melting process were used to create training data such as spatial distribution of snow depth, which has no past actual measurement values, and snow cover extent. The model trained by the data for two years during winter season were able to estimate the tendency of changes in snow depth during the snow accumulation and snowmelt seasons of another year. From the comparison of test cases with different training target, it was confirmed that the snow depth range of the training target affects the tendency inferenced, and adding various data to the training dataset leads to improvement in accuracy.

In recent years, abnormal rainfall has occurred frequently, and the ocurrence of natural disasters due to debris flows is increasing.NEXCO has taken measures against debris flow as a priority issue, however the damage situation and factors are diversifying.Debris flow disasters have also occurred in small mountain streams with a small basin area.Since small mountain streams cover a wide area along the main line, it is difficult for specialist engineers to conduct field survey and identify all dangerous areas.Therefore, in this paper, we investigated a technique for narrowing down dangerous areas focusing on the topographical features of debris flow disaster areas in small mountain streams that occurred in July 2020.

We narrowed down dangerous areas by evaluating the similarity of topography using AI.In this work, 3D topographical datas measured by aeronautical radar were used as images.Furthermore, the adaptability of the similarity evaluation was confirmed by field survey, and we have improved the accuracy of AI discrimination by reflecting the results of the survey.

Point clouds are increasingly being gathered to improve the efficiency of inspecting infrastructures. In order to utilize the point clouds, segmentation is required to classify the point clouds by member. The purpose of this study was to develop a technique for automatic segmentation of bridge point clouds. 3D bridge point clouds were converted to 2D images, and then member estimation was performed by image-based semantic segmentation using Deeplabv3+, and the estimation results were reflected on the original 3D point clouds.

In recent years, the use of 3D models is expected in the maintenance and management of bridges, and the automatic generation of 3D models is considered to be beneficial. We conducted a basic investigation of a method for automatically extracting individual planar regions representing the surface of bridges from point clouds. In this study, planes which form the surface of the structure were extracted using an improved method of RANSAC, and the point clouds on each extracted plane were selected by DBSCAN. The intersections of the extracted planes were used to capture the edges of the structure, and the planar graph method was used to construct the surface of the structure.

In this paper, we discuss the possibility of using Artificial Intelligence (AI) in infrastructure management, focusing on the analytical performance and interpretability of models. In particular, the paper outlines the mathematical background of ensemble learning methods, such as XGBoost, LightGBM, CatBoost, RandomForest, and decision tree analysis, which have recently achieved good results in machine learning applications. We report on the results of trial estimations of bridge deterioration determined using these methods. In addition, this paper discusses the analysis results from the viewpoint of AI application in infrastructure management, considering the characteristics of each method.