Recently, extending the service life of the bridges has come to be studied. Many bridges are built in the period of high economic growth, and they reached their service life. Consideration of rebuilding and extending the service life must be carried out.. Yokohama city conduct regular inspection of the bridges every 5 years. damage of the bridge can be assessed by regular inspection. Many bridges suffer from earthquake vibration and disturb lifesaving and the supplies transportation. Yokohama city carried out the seismic strengthening of the bridge. However, according to the periodic inspection, the damage degree of each bridge has unevenness. There is possibility that the bridge which the existing damage increase the risk for suffering from earthquake. In this study, deterioration of bearing which is subject to earthquake was predicted statistically using bridges inspection data in Yokohama city. Using the damage degree that we predicted, seismic risk was evaluated.

In Japan, road disaster prevention inspections began in the wake of the Hidagawa bus accident in 1978, and a great deal of technical knowledge has been accumulated to date. It is desirable to implement this knowledge in developing countries. Although the risk assessment system in Japan has a quantitative scoring system, it is assumed that an evaluation based on qualitative comprehensive judgment by a professional engineer is necessary in practice. Therefore, to expand the program overseas without specialized technology, the authors have proposed a risk evaluation system. This system uses the risk evaluation results accumulated thus far to enhance the current evaluation system and enable evaluation of risks using AI.

In Japan, slope failures associated with earthquakes and heavy rainfall occur frequently. In order to assess the damage, several organizations, including the Geographical Survey Institute (GSI), have drawn maps showing the slope failure area from aerial photographs. However, in the mapping process, the workers are manually reading the slope failure area visually, which requires a lot of labor and costs. In addition, it is difficult to map the area manually, which hinders the rapid assessment of the damage. To solve this problem, research is being conducted to detect slope failure area using artificial intelligence technology such as deep learning. In this study, we propose a method for automatic detection of slope failure regions using semantic segmentation by deep learning. The establishment of this method is aimed at efficient detection of slope failure areas and rapid assessment of damage.

This paper describes results of centrifuge tests aiming at modeling subsidence of levee slope surface where backward erosion piping undergoes during flooding events. The shape and size of the subsided trough is successfully correlated with location and size of piping in the levee or foundation soil. This model is applied to predict the location of piping in a levee where sand volcanos appeared during recent flooding events. DEM of the levee was obtained and troughs on the slope were detected near the sand volcanos. The predicted pipes were confirmed by penetration tests which directly detected loosend soil locations.

In recent years, unpredictable slope failures caused by extreme weather have been increasing in Japan. Management companies are required to detect the slope failure quickly to ensure the safety of expressway users. In order to solve this problem, a new change point detection system is proposed to detect unpredictable slope deformation by using Change Finder with SDAR algorithm for RTK-GNSS data. However, there are still problems with detection accuracy. In this research, we attempted to improve the detection accuracy of Change Finder by preprocessing multipath errors and random errors included in the data of satellite positioning system, RTK-GNSS. Specifically, sidereal time-based differential method was applied to the processing of multipath errors, Low path filter was applied to the processing of random errors. As a result, the detection probability by Change Finder was significantly improved.

It is important to identify the model that can simulate the field measurement data of soil moisture conditions to predict the occurrence of landslide disasters due to heavy rain. This study verified the applicability of the recurrent neural network to predict the field measurement data of volumetric water content. The recurrent neural network model was estimated by using the training data at the time of weak rain, and the estimated model simulated the test data at the time of heavy rain with enough accuracy. The simulation results led to the conclusion that the recurrent neural network was an effective method to predict the field measurement data of volumetric water content.

The purpose of this study is to improve the accuracy of water level prediction several hours ahead in order to carry out more appropriate flood control activities and evacuation behaviors against flood disaster.

The target of prediction was the Hiwatashi basin of the Oyodo River system, and the water level prediction was carried out by using the information of the upstream water level observatories and the rainfall observatories around the basin. As the water level prediction model, GBDT (Gradient Boosting Decision Tree) model was used to predict for 1 to 6 hours. Prediction of stacking model is performed stepwise as follows; First, a short time prediction is carried out. Then the prediction result is added as a feature quantity for the next time prediction. By repeating the same procedure, we made prediction model up to 6 hours ahead. We compared the accuracy of the constructed model with the model without stacking, and confirmed the accuracy improvement by the stacking model.

The effect of the inflow prediction error was investigated in a dam operation model, which is based on the neural network with reinforcement learning. The dam operation model used in the study determines the appropriate dam discharge according to the situation that changes from moment to moment (reservoir water level up to the present time, observed inflow and timeseries of predicted inflow up to 6 hours later). The examination object was Matsubara Dam of the Chikugo River system. For model training, virtual flood data created by extending historical rainfall was used, and reinforcement learning (Deep Q Learning) was applied. In order to verify the model, artificial error was added to the inflow of virtual dam inflow data, and used as the virtual inflow prediction data during actual operation. In the verification with three virtual floods exceeding the design flood scale, the model showed a reasonable operational judgment and hardly affected by the error of the predicted inflow.

Among the deterioration of 700,000 bridges in Japan, making the bridge inspection once every five years sustainable is a fundamental issue. In order to speed up the cycle from inspection to repair measures, it is required to inspect consistently health condition while suppressing variations, and to select a repair method using numerical indicators. In order to extract the pixel-wise region of bridge elements, we propose a method that learns a target region detector useful for feature extraction of damage by semantic segmentation using the dataset for the images of the floor slab and its crack annotated labels. This method automatically calculates a deterioration index for scaling crack area．Finally, we address the issue of generalization of bridge inspection support tools.

It is obligatory to carry out regular visual inspections once every 5 years for aging road bridges. However, it is difficult to carry out inspections with the same quality in the future due to the problems of lack of financial resources, human resources, and technical capabilities in the inspection by visual inspection. Under such circumstances, it is expected that the inspection work will be made more efficient by the automatic damage detection technology using images. In this research, we develop a technique to assist crack detection, which is one of the inspection items, by using image recognition. The images of a wide range of concrete structures taken by a high-resolution camera were mesh-divided into multiple images, and the presence or absence of cracks was determined by an image classifier created by CNN. In addition, by combining the results of moving the positions of mesh division and taking the average value (Average Shift Mesh), the detection accuracy of the cracked portion and the resolution of the output result were improved.

In maintaining and managing social infrastructure, saving labor and improving efficiency in visual inspections is one of the urgent issues in the recent years. Therefore, in this study, in order to solve this problem, we apply segmentation method using deep learning to images acquired by a digital cameras and examined a method that automatically detects the deformation and classifies the degree of damage. In addition, we also investigated a method that uses depth images, and confirmed that the RGB-D camera is an effective device for labor saving and efficiency of visual inspection.

When creating a 3D model of a bridge for maintenance purposes, if the location of damage is reflected in the model, it is expected that the location and shape of the damage can be easily confirmed. Furthermore, cracks are the most common damage to bridges. Although various methods for the automatic detection of cracks from images of concrete surfaces have been studied. These methods make it difficult to detect cracks when objects other than concrete appear in the image. In this study, Semantic Segmentation is used to extract only the concrete area from the bridge images, which enables the detection of cracks from various images taken. In addition, a 3D model of the bridge was constructed by Structure from Motion using the images in which cracks were detected, and a 3D model reflecting the location of the cracks was constructed.

In order to accurately evaluate the performance of existing concrete structures, a method that can acquire not only the surface but also detailed information such as crack positions, sizes, and angles inside the structure is desired. Therefore, in this research, we attempted to apply the method of estimating and visualizing the position and angle information of defects in concrete using GAN from radar images. Specifically, first, a specimen with artificial defects arranged at different positions, sizes, and angles is prepared, and radar images are acquired. The acquired data was learned by pix2pix, which is an application technology of GAN, and the cross-sectional image including defects was estimated and visualized from the radar image. As a result of the examination, although the proposed method can reproduce the position, size, and angle information of the defect, when the position of the defect becomes deep, the reflected wave from the defect becomes relatively close, and as a result, the estimation accuracy tends to decrease.

Current maintenance methods of revetments are inefficient because it depends on the visual inspection of civil engineers. Thus, it is necessary to develop a quantitative and rational inspection method.

In previous studies, the authors developed a crack detector using CNN. However, in order to utilize detecting results in the current maintenance standards, crack width must be measured in actual size. Therefore, this study considered a method for measuring cracks from detecting results. As a result, the method is developed that can measure width of the entire crack.

It is a popular to hammer sound test for visual inspection of deterioration in concrete structures. This convenient method is much effectiveness for expert engineers. However, it is difficult for young engineers and applicable to robotization to apply quantifying and to be systematic under consideraton to complicapable of relation of sound data and degree of deterioration. The factor of relation to hammering sound data and degree of degradation are not clearly. Development of quantifying and being systematic as engineering are the most important in practical business. In this study, it is expected to apply using AI technology. In this study AI is expressed by machine learning, in particular deep learning based on neural network. It makes a clearly for effectiveness of methods of machine learning and propose to apply to autoencoder.

Impact-Echo monitoring is widely used as a method to find the flaking or cracking caused by deterioration of concrete structures. However, the accuracy of the monitoring depends on the skill and experience of the engineer. In this paper, we investigated the possibility of machine learning to quantify the Impact-Echo monitoring. Specifically, the impact sounds were imaged by spectrogram conversion and classified by a convolutional neural network. This method was tested on concrete test specimens with simulated defects and also salt-damaged real structures. It was possible to diagnose concrete defects with the same accuracy as a skilled engineer by using small image data. It was shown that the obtained neural network has a certain generalization performance.

This paper attempted to detect initial damage of steel finger type expansion joints based on passing sound measured in a vehicle. Firstly, it has been clarified the feature of passing sounds over test specimens which are simulated the fracture joints. Using the feature of passing sounds over test specimens, damage evaluation of actual expansion joints were possible. Moreover, the measurement were conduted one year later to collect and compare the feature of passing sounds. Consequently, by collecting the passing sound until the expnansion joints damage, the possibility of damage detection by measured data from inside the vehicle was verified.

Track management is important to improve the safety and ride comfort of trains. The unsupported sleeper, which is one of the track anomalies, is currently difficult to be detected by conventional track inspection methods, because of its short wavelength components and appearance only under train passage. This study proposed the method for detecting the unsupported sleeper from vehicular vibration responses, based on highly nonlinear time series prediction by LSTM and the estimation of uncertainty by Monte Carlo Dropout. A multibody dynamic model of the express train was utilized, and the effects of the unsupported sleepers on the dynamic performance of the train were studied. The proposed method was then applied to the simulation data and its performance was evaluated. It was confirmed that the the proposed method improved the detection ability comparing with conventional methods, while obtained high accurate detection for the cases with two or more unsupported sleepers.

The deterioration of bridges has become an urgent problem. Bridges need to be properly maintained and managed, so that bridge managers have been required to perform close visual inspection once every five years since 2014. However, the continuous close visual inspection is difficult in the local government in which finance and manpower are insufficient. Therefore, alternative means of close visual inspection are studied. The authors developed "Bridge inspection support system" which detects damage from bridge images by AI analysis. The inspection by the image can be expected for efficiency improvement and labor saving. On the other hand, in order to upload bridge images with a huge amount of data quickly, high-speed communication is necessary. In this study, the bridge image was uploaded to this system using 5G, and we examined the usefulness of the remote bridge inspection while communicating with the bridge site in the remote place. We experimented on the uploading time of the image in the simulated remote bridge inspection using 5G. In addition, we interviewed bridge inspectors about the possibility and problems of the system, and grasped the problems of communication in remote bridge inspection

With the progress of sensors and communication devices, various infrastructure data can be acquired by wireless connection. The data acquisition method using LPWA communication is highly anticipated because it does not require a radio communication license for installation, low power consumption and a large number of sensors can be installed at low cost. However, as the data characteristics of infrastructure, data during a disaster is the most needed, and continuous data collection and distribution is required even in the extreme situation of power loss. The authors are studying to realize a system that enables continuous monitoring status of multiple sluice gates even in a critical situation such as public communication disruption due to a long blackout caused by a huge typhoon. Including satellite utilization, enhancement of resilience of LPWA network data acquisition at power loss is discussed.

The Number of cruise ship calls has reached a historical high in Japan on 2018, and the cruise tourism became one of important subject to grow Japanese tourism industry. However, little study has been done to understanding the travel behavior of cruise tourists. In this study, Wi-Fi packet sensing are used to observe travel behavior of cruise tourists. It became clear that differences of travel behavior of cruise tourists may exists through an analysis of difference in visit rate of each travel spot using the data observed by Wi-Fi packet sensors that placed in major travel spot in Kanazawa City.

In recent years, with the declining population and the aging of society advancing, there is a possibility that problems such as a decline in living convenience and a loss of appeal of town due to diffusion type urban structure may be encountered. Therefore, in order to aim for an intensive city, it is important to revitalize the central urban area with stocks of urban functions and infrastructure to some extent.

In this research, in order to analyze the current state of the central urban area, in order to grasp detailed citizen behaviors of citizens, from the large questionnaire survey, the characteristics of the inhabitants inside and outside the city center, the average number of roundabouts in the urban area , Behavioral characteristics by day and night, dissatisfaction degree which is a factor that hinders the visiting town. In addition, in order to grasp the trend of citizens who can not obtain by questionnaire survey, survey of Wi-Fi packet sensor in downtown at night is conducted using raspberry pie of small computer and the time zone distribution of people is shown for each area.

Japan has many rivers, and the almost management work is carried out manually. Due to the development of drone (UAV) technology, they are applied to river monitoring, and the research with AI, such as using the image taken by drone for detect illegal dumping, is also increasing. On the “ground”, taking a lot of picture for report by humans requires hard work and is also subject to the diversity of condition. The aerial image has a different angle of view from the ground image and the conditions are different. In this research, whether can utilize the ground image of river maintenance management database system RiMaDIS to improve the detection accuracy in less aerial image are verified. The image of illegal dumping taken from ground and drone are classified by three image feature value, and they were learned by Faster R-CNN. The verification result suggests that features value such as Bounding Box occupancy contribute to the improvement of detection accuracy.

Japan has many rivers in its land, and many river maintenance operations such as inspections and patrols are performed by visual inspection. Especially in frequent river patrols, personnel record anomalies and grasp the condition of river channels on site, while some of them can be made more efficient by combining UAV photogrammetry and image recognition by deep learning. In this study, we propose a method to classify regions such as sandbars and trees in river channels by applying Semantic Segmentation, which is one of the image recognition methods by deep learning, to ortho images obtained by UAV photogrammetry. The deep learning model was trained using aerial images with different shooting altitudes, and a highly accurate river region classification model could be obtained. We conducted an experiment to apply the model to ortho images with different ground sampling distance (GSD), and confirmed that general region classification can be performed even for ortho images with different GSD from training data.

UAV based bridge patrol can roughly scanning the surface of the bridge and bring back high-resolution image data in very low cost. To use those data smartly, the application of Convolutional Neural Network (Deep Net) based image processing can found deteriorations such as cracks, palling, corrosions, leaking water effectively. Efforts have been conducted to extract labeled damage photos from past Inspection reports and to find out effective way to train CNN model for damage detection. Although it can achieve accuracy about 90% after data argumentation using this data base, its ability of recognize damage from real word UAV image is low. In this study, mixed trainings of CNN model with both UAV sourced image data and inspection report sourced data were conducted to reinforce the machines performance when it’s seeing the background and no-damage structural members, which are less in the inspection reports but resourceful in real UAV scanning images. UAV videos acquired from a few real bridge patrols were used as sources. The 4K images was sliced from video and split to small samples. Each sample was then labelled manually to classes including background, a few types of damage, and a few types of undamaged structural surfaces. Training with original inspection report sourced data, UAV sourced data, and mixed data were conducted and compared in accuracy of damage recognition in UAV image.The data set is used to train a Fully Convolutional Network (FCN) model for detecting damage image and location on small pieces of images. And the location of damage and its category can be shown in a stitched image visualizing the damage and its distribution.

A hydroelectric power generates an eroctronicity with potential energy of water flow. The water is carried from fountainhead to hydroelectric plant path through a penstock. A penstock diameter is approximately tens of centimeters to several meters, length is approximately tens of meters to several kilometers. A regular maintenance method of penstock needs worker to approaching penstock and check it surface by their eye. Maintenance worker have to walk the distance from one end to the other of penstock to check the whole. This method needs much time and cost.

We proposed automatic maintenance method using a drone and image processing method and evaluated it. On the other hand, an penstock is located at the outside in many case. Therefore, an penstock is sometimes covered by mad, leaves, glass and so on. It is difficult to distinguish such noises by visual image processing method. We focus on an infrared image to improve proposed method. We discussed a probability of infrared image by compared infrared images and visual images taken by drone.

For efficient dam body surface inspection, we propose the UAV autonomous navigation in a non-GNSS environment using a total station, the accuracy improvement method of photogrammetry technology using images with geotag by autonomous navigation, and the aging detection of the dam body surface using the CycleGAN.

In addition, we implement a field study to demonstrate this proposed method, and show its usefulness.

Inspection of infrastructure is regarded as the most important issue in order to safely maintain and manage infrastructure facilities including aging bridges. However, the number of engineers who can inspect bridges is currently under serious labor shortage. While efficient maintenance of bridges is an issue, small unmanned aerial vehicles, commonly known as UAVs, are attracting attention. It is thought that the stability of operation mentioned in the problem of bridge inspection using UAV can be improved by using autonomous flight UAV. In this study, flight test and thrust test were performed to evaluate the performance of a general-purpose UAV in bridge inspection. In addition, Waypoints were created using a smartphone application. Using a general-purpose UAV that had a high Performance evaluation value in bridge inspection, an actual bridge inspection by an autonomous flight UAV was performed. 2D models of the inspection bridge were created to automatically detect damage from the video taken during the actual bridge inspection. Finally, as a future perspective, the necessary items for a complete UAV autonomous flight inspection of the bridge were proposed.