Introduction of Artificial Inteligence into the field for maintenance of infrastructure is now inevitable and urged in order to solve the serious problem such as short of young generation who sustain the infra-structure in future. The Public Works Research Institute, to which the auther belongs,began to engage the project to apply AIs to road bridge maintenance in this context. This occasion makes us to consider the meaning of sytstemize the maintenance prosess of road bridges. This paper describes current progress of the project though it is still on the way. We believe that this inform some important facts that we have noticed to the people who are making similar challenges.

In recent years, as the number of workers in the construction industry has been declining, the use of artificial intelligence technology has been a major issue to improve the productivity. In particular, the development of supervised machine learning technologies, such as deep learning technologies, has been remarkable, therefore it is expected to be used in the field of civil engineering. However, it is hard to say that the practical application of AI technology has made steady progress compared to other fields. In order to provide direction for future work by researchers in the field of civil engineering, this paper provides perspectives on data platforms for data integration and storage, measurement automation for data collection, and methods for linking data and knowledge.

In recent years, the effective and highly advanced maintenance has been implemented utilizing ICT and AI in the field of infrastructure maintenance. In particular, i-Construction aims to improve productivity by employing ICT in all phases from survey/design to construction and maintenance. Therefore, it is required to built a data platform that integrates and centralizes various data of each phase seamlessly.

This paper describes the infrastructure data platform (i-DREAMs^®) implemented on the Metropolitan Expressway. In addition, the affinity of the platform with CIM, evolution to “a comprehensive disaster prevention information system” that integrates information in urgent times of disaster, and the extensibility to other systems are described.

The objective of this paper is to present the process of data science in the field of infrastructure management. In this paper, we describe the method of analysis, the viewpoint of interpreting analysis results, and the viewpoint of evaluating the business improvement effect, which are essential for every data science process. We also discuss a list of cases in which the results of data science processes are utilized for infrastructural management operations. In addition, we describe the system architecture, pre-processing of data, and inquiry processing while using data on the proposed data-utilization platform that can rapidly process various types of data. Furthermore, in this paper, we discuss aspects such as accelerating data visualization, analysis, and data science trials by employing the proposed data-utilization platform.

The true value of civil engineering is building technical systems from limited information subsets and offering enough efficient facilities for societies. But their selective information is easy to reach tascit knowledge and is large burdens for information share. On the other hand, we think that Deep Learning technologies can extract much practical information which we haven’t assumed and realize work efficiency and productivity improvement by formal knowledge. But when their technologies are directly applied to human oriented original flows, they do not often reach human levels. In this article, we describe essences that we apply Deep Learning technologies to infrastructure maintenances.

It is believed that the introduction of AI will cause changes in work and work styles. For example, it is said that inspections in factories will be replaced by using image recognition technology, and that call center operations such as Q&A will be made more efficient by using chatbot-like support. When it comes to maintenance of civil engineering structures, it is often said that visual inspections are replaced by drones. It can be said that these issues are how to improve productivity by introducing AI technology. However, it is assumed that the introduction of AI will not only replace the existing work, but also that the changes will be diverse such as "work", "work range", "work efficiency/productivity", and "motivation for work". In this research, the concept of how infrastructure maintenance work is reformed by the introduction of AI is summarized.

This paper describes a societal implementation of national projects by our two research projects: “R&D of Laser Directive Noncontact Diagnosis System for Maintaining Degraded Infrastructures” and “Development on skill transfer of professional engineers and advanced scanning devices in infrastructure management”, parts of Cross-ministerial Innovation Promotion Program (SIP) managed by the Cabinet office of Japan. Main challenge of societal implementation of these national projects is to commercialize project achievements by research institutions with insufficient experience of business planning. Through our several years activities in a startup company including analyses of the target business field and the owned technology, we found that business strategy is important to establish equal partnership with the existing major companies for market penetration, and neutral business planners play an important role to coordinate all institutes’ interests.

In the expressway which is important society infrastructure, the stock is enormous, and the productivity improvement of the periodic inspection becomes the problem. Therefore, many studies for a diagnosis of the structural soundness with AI technology are done. For the further productivity improvement , it is effective that the structural soundness can be set the priority of measures. As one idea, there is a method to extract the percence of influence by injury and initial defect by comparing the degree of soundness decreased due to the deterioration, injury, and initial defect with the degree of soundness decreased due to just the deterioration.

Gradient Boosting Decision Tree (GBDT), a machine learning technique, is often used in data analysis competitions because of its superior accuracy and computational speed. The purpose of this study is to investigate the applicability of GBDT for estimating the cause of damage and the repair method from the information of bridge management records stored in a bridge management system, in order to assist the decision-making of road administrators in local governments in selecting the repair method.

As a result, GBDT is capable of estimating the cause of damage with high accuracy for all models. In addition, the influence of the specifications on the cause of damage is analyzed by using importance and SHAP values of the explanatory variables.

A large number of inspection records have been compiled to date as records of bridge inspections. However, most of them are in paper form or the data is saved in PDF format, which makes it difficult to utilize the accumulated data. If it is possible to build a database of inspection records by automatically extracting data from inspection records, it will be possible to analyze data based on a huge number of inspection records. Moreover, the element numbers, which are important information to know the location of damaged members, etc., can be found in the bridge drawing. However, a lot of the element numbers overlap with lines of the structural members, making it difficult to extract them by general OCR processing. In this study, the element numbers listed on the bridge drawings in the inspection record were extracted utilizing object detection by deep learning. As a result, it was confirmed that the extraction of the element numbers that overlapped with the lines was possible with high accuracy.

In Yamaguchi Prefecture, the inspection data is expected to increase due to the periodic inspection of road facilities, and the efficiency of data management is required. In addition, due to the electronic administration open data strategy formulated by the government of Japan, it is obliged to take measures to promote the use of public data held by the national and local governments, and to make the people easily access the public data via the ways like the Internet. However, the existing inspection data cannot be directly opened and utilized. In this study, an inspection file management system for the data of periodic bridge inspection in Yamaguchi Prefecture is developed, and the importance of data cleaning in data opening is clarified due to the existing of files with incorrect names and missing files. In addition, a API system for open data is developed in this paper.

Application of deep learning methods to bridge inspection has been studied in order to resolve shortage of engineers and increasing cost for periodic bridge inspection. To improve the efficiency in bridge inspec-tion, AI model should support various damage types and assessment criteria. This study aims to propose a multi-task learning method related with damage classification for high-accuracy and robust damage assessment model. Normal multi-task learning and multi-task learning with attention mechanism relating damage classification to damage assessment explicitly are validated through the comparison with single-task learning method using bridge inspection data in Nagasaki Prefecture.

In the periodic bridge inspection in Japan, engineers are required to determine the damage level for each structural member by visual inspection in all of seven hundred thousand bridges. More sustainable and low cost bridge inspection system is required in the future. For this purpose in this study, the applicability of deep machene learning to the damage level determination of bridge members is verified based on the visualization technique. In detail, the convolutional neural network (CNN) that determines the damage level of the bridge member from the image data was constructed. And then, the Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to verify the features of image, which contribute on the damage level determination. By comparing the heat-maps of the Grad-CAM, the consistency of the feature to determine the damage level in the CNN to the feature used in the inspection conducted by expert engineers could be discussed. Furthermore, it was also shown that knowledge from the outputs of Grad-CAM is applicable to the improvement and encourage of acceptance of the CNN.

In Japan, there is a discussion about the replacement or extension of the service life of the bridges built during the rapid economic growth period, which are now reaching the end of their planned service period. As issues with the continuing close visual inspection of bridges are surfacing, the remote imaging system is expected to become a new inspection method that replaces close visual inspection. Although the automation of the creation of the data of damage has been achieved, the automation of the diagnosis of the soundness level has not been performed. In this research, we proposed a method to extract the feature quantity of damage from the damage map in the past inspection result by using image analysis such as pattern matching. In addition, the damage affecting the diagnosis was clarified by performing the decision tree analysis with the extracted feature quantity of the damage as the explanatory variable and the diagnosis result of the maintenance strategy as the objective variable.

Lean duplex stainless steel (LDSS), which is expected to apply to infrastructures, exhibits rounded shape of stress-strain curve. For this reason, a constitutive equation which is able to accurately express the curve is required for the ultimate strength analysis of LDSS structures. Authours have already proposed MRO curve as this kind of equation. However, not only 0.2% proof stress and tensile strength, which are specified in common material standard and a mill certificate, but also mechanical properties such as proportion limit etc are needed to describe the equation. In this study, we collected tension coupon test results of LDSS and created the simple estimated equation by means of linear regression analysis. Also, we predicted the mechanical properties by using Random Forest (RF) which is one of machine learning method. According to comparison predicted results by RF with those by estimated equation, it was revealed that RF has same prediction accuracy of mechanical properties as estimation equation.

Corrosion crack width can provide effective information to evaluate the deterioration level of in situ corroded RC structures. However, uncertainty associated with the relationship between corrosion crack width and steel corrosion in the RC structures is quite large. In this study, a probabilistic framework for estimating residual strength of corroded RC structures using the observed corrosion crack width distribution and Long Short-Term Memory (LSTM) network was proposed. In an illustrative example, effect of corrosion crack width on probabilistic density function associated with flexural capacity of existing RC beam was investigated.

In this study, the concrete shrinkage and creep laboraty data is analyzed based on the regression by machine learning, linear regression and the design empirical predictionin Japan. The random forect predicted the ultimate shrinkage under various conditions most accurately, while the ultimate creep was estimiated by the neural network with maximum accuracy. It was found that the machine learning can approximately predict shrinkage and creep under conditions beyond the design range but is not able to estimate them under extreme conditions such very high relative humidiy close to 100%, high water-to-cement ratio over 0.8 and others The importance of parameters according to the randam forest was reasonable to reflect shrinkage and creep characteristics known by laboratory test and design. The machine learning based on the laboratory experiment cannot reasonably predict the variation of shrinkage and creep whose learning data is few and the extrapolating long-term behavior.

Despite the fact that harbor structures are extremely vulnerable towards chloride induced deterioration, study on residual structural performance of corroded steel pipe pile pier is still limited. Authors have previously proposed an evaluation method of residual structural performance based on the judgement result of deterioration by versatile structural analysis software. However, a simpler method without conducting structural analysis is appreciated when deciding the maintenance priorities of different piers. In this study, authors proposed evaluation methods of residual structural performance by introducing artificial intelligence to estimate both the area ratio and distribution image of damaged beams. In order to overcome the problem of misrecognition in unlearned piers’ shapes, a new idea of image processing is hatched by considering the relative distance of beams. As a result, the shapes of unlearned piers are correctly recognized and the estimate accuracy for damaged beams are confirmed.

Interstitial pneumonia causes abnormal iflammations of lung tissues and those are represented in CT images. Therefore, state of the disease is judged from the CT images manually by medical specialists and computer-aided diagnosis is expected. In this paper, we develop an image processing system with machine learning techniques in order to support the diagnosis. At first, we construct an image processing to extract lung regions from CT images by using genetic programming techniques. Then, an index is proposed to identify patients of interstitial pneumonia from information of the extracted lung regions. Finally, each CT image is classified into 6 patterns according to diseased area in the lung regions with the aid of convolutional neural network. Consequently, the network can classify the CT images with more than 90% accuracy.

In Japan, technological development is being carried out using various ICT, IoT, AI, etc. in order to cope with a huge number of bridge inspections. In this study, we acquired the surface condition using LiDAR for the purpose of grasping the surface condition without contacting the infrastructure structures such as bridges. In addition, we conducted basic experiments on human tactile sensation in subjects. As a result of the correlation analysis using LiDAR data and human tactile data, it became clear that it is possible to grasp the "rattling" feeling without touching the object. On the other hand, it has become clear that it is necessary to use a higher-performance sensor for delicate irregularities on the surface of objects such as "smoothness" and so on.

This study aimed at the traffic sensing based on B-WIM algorithm that analyzes the traffic environment using the structural response of bridges. The time-frequency analysis based on the wavelet scattering transform extracted features from traffic-induced vibration of a bridge in service. The wavelet scattering transform calculated the scattering coefficients through the multi-layered convolution of input signals by the wavelet and scaling functions. To learn neural networks from limited number of traffic data, the learning data was amplified by subsampling the scattering coefficient in the time direction. An identification flow of traffic vehicles was proposed based on multi-level classification with learned neural networks. Even from a small number of traffic dataset, the identification of a test truck and local buses was achieved with high accuracy by amplifying the learning data.

In the asphalt pavement inspection, efficient and quantitative methods are required. So in recent years, several methods for automatically detecting cracks from pavement images has been proposed and its effectiveness has been shown. This paper have attempted to construct a system to efficiently improve the accuracy of crack detection by convolutional neural network which is a type of deep learning. It is generally considered that the accuracy of deep learning model is proportional to the amount of learning data. Therefore this paper have not simply increased the amount of learning data, but used only the data that satisfied the conditions. When the accuracy of the model obtained by this method has compared with that of the model learned from randomly selected data, it has confirmed that the accuracy of the model by this paper method has stably improved. In addition, the soundness judgment of pavement was carried out using the crack detection result of the model, and the result was mapped on the GIS.

In current, in order to realize a sustainable society, the maintenance of road and other infrastractures in Japan tasks are considered important. However, reduction of public works expenditure due to administrative reforms and the shortage of civil engineer due to the low birthrate and the aging population. In particular, we have been focusing on the evaluation of road pavement condition of local goverments and have been working on the development of a simple assessment system. And, development of AI technology in recent years, an efficient system that replaces conventional operations is also expected.

In this study, purpose to improve the efficiency of road administrators in local governments, using deep learning we attempted to automatically detect the deformation that affects roughness from the video data shooting by digital camera. And, road pavement condition system using acceleration sensor and video data developed in the past study, the usefuluness of the road evaluation by the application of the deformation detect is confirmed.

To predict change in road using construction works data helps mapmakes to update road maps. This paper describes the result of experiment on predicting road change using document vectors by title, items list and material list of construction works data to evaluate their effectiveness. This paper also reports comparison of their token by different extracting method. According to the results, document vectors by construction works were useful to predict road change. It’s accuracy was 0.83 and recall was 0.85 using character 2-gram of construction title through decision tree. Despite title and item list were less amount of data than material list, the former showed better results than the later. In the comparison of token, short character N-gram showed better accuracy than phrases which were extracted by delimitter, words extracted by morphological analysis and dictionary, word N-gram and clustered phrases based on word embedding.

In order to ensure smooth traffic and safety on highways in winter, we are implementing winter tire regulations.Under this regulation, tire check staff will stop passing vehicles and visually check the type. Vehicles without winter tires are guided outside the main line. However, there are various problems. For example, securing many personnel such as tire check personnel and vehicle guides, the increase of human burden due to long-term restraint, occurrence of traffic congestion during times of heavy traffic and so on.

Since 2016, the authors have begun to develop an automatic winter tire discrimination system and have been developing image processing and discrimination technology.

In recent years, even in areas where there is usually little snowfall, large-scale retention on roads is observed when snowfall occurs. The monitoring of abnormal situations by road administrators and the interpretation of road surface conditions are mainly performed visually, and the efficiency of abnormality detection is a little bit low.

In this study, as a support tool for road administrators to quickly detect anomalies and make processing decisions. We developed an AI model that automatically determines the road surface condition. The training data were made by the image of the dashcam data, which is classified into 5 types, such as dry, wet, flood, wet snow, and consolidation. As a result of automatically discriminating 26199 road surface images of day and night using the AI model, the Training Accuracy rate was around 85%.

It is important to evaluate the snow cover condition to manage the road condition such as snow removal on the roof of a building or on the road. In particular, evaluation of snow depth requires the use of expensive sensors such as lasers in addition to visual evaluation. Image processing techniques such as deep learning have improved in recent years, and many studies have been conducted to evaluate the snow cover state using images from surveillance cameras. In particular, in the surveillance images of the road surface and shoulders by wayside cameras, the location information is clear because the shooting location is fixed, and the angle of view changes relatively little. In this research, an AI method was applied to the evaluation of the snow depth of the snow on the shoulder using a surveillance camera.

Though the situation of air pollution in Japan is improving in recent years, the points or limited aera with high polluted density still exist due to heavy source or enclosed space. As a result, the complaint about air pollution occurs often. In order to improve the situation of the high density points, alerts are sent out on the website or on the local electronic bulletin board to urge the drivers to refrain their car-use. However, the alert is only based on current observed values, there is no method to predict the alert in advance now.

In this study, Street Canyon by deep neural networks is used to the predict of air pollution, and validity of the predictions are examined, then the necessary input items are discussed. As a result, it was found that each input item (temperature, wind speed and direction, and periodicity of traffic volume) was associated with NO2 and the fact that broad-based input items are required for SPM predictions. It is also shown that deep learning is useful in studying air pollution predictions under Street Canyon.

Kanazawa Port was chosen as the base port on the Sea of Japan side in 2011. Since then, the number of cruise ships calling at the port has increased. Due to the increase in the number of port calls, many tourists are visiting Kanazawa. On the other hand, around Kanazawa Port, trailer bus for the sightseeing and vehicles such as spectators are concentrated. Therefore, traffic congestion is a concern.

The purpose of this research is to quantitatively understand the influence on road traffic. Therefore, the time reliability index was calculated using the route information and travel time information obtained from the ETC2.0 probe information. The time reliability index was calculated in three cases: the day of the holiday, morning on weekdays, and evening on weekdays. As a result, in all cases, there was a section with reduced time reliability on the day when the cruise ship called. Especially on weekday evenings, the time reliability was significantly lower than in other cases. It was revealed that the call at a port of the Cruise ship had a big influence on traffic of the evening on weekdays.

We propose using Centroid tracking for traffic counting. We use normal video camera instead of specially design for traffic counting system. We also design dataset for traffic counting by mixture picture for each category. We also en-hance COCO dataset tools to handle our dataset. Using mixture category dataset we can improve detection.

In recent years, machine learning methods such as deep learning have evolved greatly in terms of performance, and they have been used for various purposes in disaster prevention. On the other hand, the intrinsic shortage of the number of data, the improvement of explanatory and interpretive nature of the task processing process are important issues that need to be addressed by computational models for decision making in disaster management. In this paper, we will discuss the concept, methods, and applications for addressing these two points based on research trends in the field of machine learning, and also introduce approaches that integrate mathematical and data-driven models to address these two issues.

Hanshin Expressway is conducting research for next-stage infrastructure management that makes optimal decisions using various data in the real world such as road structures, vehicle traffic, energy, and the environment. In cyber space, we are utilizing the accumulated data and working on simulations using a digital twin model.

This paper introduces the concept of "Hanshin Expressway Infrastructure - Cyber Management System", its efforts and examples of future visions.

In recent years, it is expected to proactively introduce AI technology in the field of civil engineering. However, lack of knowledge related to AI technology makes it difficult to properly grasp the effects of the introduction of AI technology and to accumulate various data for using AI technology, and will be an obstacle to introduce AI technology. This paper, with the goal of making teaching materials, outlines the basic knowledge necessary for understanding AI technology for civil engineers who want to utilize AI technology for various problems in actual operations. Firstly, the histroy of AI is explained in order to properly imagine the AI technology indicated by the word of AI. Then, it is organized the ideas and inputs/outputs necessary to use AI technology. Lastly, through actual cases, understanding of AI technology is expected.

Pattern recognition is a process of extracting certain features and rules from data, which is achieved with high accuracy by machine learning algorithms such as deep learning. Because the essence of deep learning is interpolation of training data in a feature space, it remains a main issue for data-driven models to obtain extrapolation ability and interpretable computation process. This paper introduces data-driven techniques for law-discovery, which build a natural computational process and obtains generalized prediction ability by discovering laws behind the data, and discuss the current and future trends of the framework.

Sewerage has been developed with the major objectives of "inundation control," "improvement of public health", and "conservation of water quality in public water areas". In recent years, in addition to these purposes, the practical application of the recovery of resources and energy of sewage such as phosphorus recovery and methane fermentation has been attempted. On the other hand, although the facilities are becoming obsolete, it is difficult to renew the facilities due to the financial situation and the declining state of the population. Therefore, it is required to appropriately manage the existing facilities. In order to achieve these objectives and perform more appropriate value-added operation management, it is necessary to collect the necessary and sufficient information and utilize it. In this study, the possibility of various information that sewerage has is considered. The water quality of effuluent from a wastewater treatment plant is predicted by machine learning using the operation & management data acquired every day as an example of information utilization. The ideal ways of sewerage as an information base in the catchment area are considered and discussed as well.

Typhoon No. 15 in the first year of Reiwa, which hit the dawn on September 9, 2019, caused a great deal of damage mainly in the southern part of Chiba prefecture. Because of the seriousness of the damage, it was designated as a severe disaster by the government on October 11, 2019, and named "2019 Boso Peninsula Typhoon" by the Japan Meteorological Agency on February 19, 2020. In the case of a disaster over a wide area, it is necessary for the local government to collect the damage situation, and for the prefecture and the country to gather information and recover. However, in Typhoon No.15, it was difficult to collect information in local governments due to the effect of large-scale and long-term power outages, which greatly affected the recovery speed. In this study, in order to prepare for a huge typhoon in the future, we investigated and analyzed the damage situation of typhoon No. 15 by integrating various kinds of information in the southern area of Chiba prefecture. The results show that there is a time lag between the time of disaster and the time of application for a certificate of damage to a house. In addition, we report on the results of predicting the congestion of application windows in city halls using deep learning that can predict time series.

A confirmation of final tightening of bolt is peformed by practicing engineers. They check a marking on the each bolt to judge a tightening correctly performed. However, this procedure has some problems. 1) The huge number of bolt requires many time to be confirmed by an engineers and 2) the objective record of the confirmation of all bolt are not remain.

We propose an automatic confirmation system of final tightening of each bolt using deep learning based image processing method. The system receives the movie of bolt tightening area as input. At first step, the system detects each bolt in the movie and clips images of one bolt. Next step, the system detects marking on the bolt from the clipped images. Last step, the system judge detected marker means correctly fasten of each bolt or not. We conducted the experiment of the system with the movie of real bolt tightening area. The results show some problem at shooting a movie in real circumstance and we discussed these problem.

At the Ministry of Land, Infrastructure and Tourism, accumulation of weight for the materials brought into the construction yard by crane is implementing as one of the methods for grasping the production amount at the construction site. In order to accumulate the amount of production as weight, three elements such as material type, weight, and transport location are necessary. In the trial construction, materials were classified by video recording, weight was measured by crane scale, and the transportation location was determined by the location information system. The transportation location is for determining whether the material is loaded into or unloaded from the construction yard, or moved within the yard, and does not require high accuracy in the measurement coordinates. This article is describing the method for grasping the approximate plane position of the crane scale in real time by using triangulation from two fixed-point camera images and crane scale data which is gained from the learning data of the object detection algorithm YOLO tried by automatic distinction of material type.

In recent years, efforts have been made to improve the productivity of construction work using information technology. However, since humans operate the construction equipment with machine control and machine guidance, manpower has not been saved. Therefore, if autonomous control of construction equipment by reinforcement learning is possible, it will be possible to reduce manpower by automatic construction. In this study, a reinforcement learning algorithm, PPO, was used to generate drilling motions by an agent assuming a excavator. As a result, the motion was successfully generated and at the same time, the speed of the motion was increased and the amount of drilling was maximized. In addition, the future prospects of civil engineering construction by reinforcement learning are described based on the findings.

The development of AI technologies that can derive information on construction management from onsite images has been developed very actively in the decade. We developed an AI-based system to detect construction machines from the site images automatically for site performance management. As a result of operating this system, the detection error was ±14%. To realize highly accurate detection, the amount of learning data for AI and the bias of the orientation of construction machines have significant influence. In this paper, we focused on the number of labels and direction of construction machines and examined the effect of learning data on detection error. As a result, we showed that it is possible to improve the detection accuracy with less learning data by creating learning data by setting a certain rule instead of increasing the learning data randomly.

Trees in urban areas play important roles such as alleviating heat island effect, absorbing carbon dioxide, and offering bio-diversity. Tree monitoring is, first of all, crucial to effectively manage and utilize the trees. Many efforts have been done to monitor the tree structure using image analysis. For example, tree species classification and above ground biomass estimation were conducted using satellite image analysis. Tree measurement was also performed using laser scanner called LiDAR mounted on airplane, helicopter and drone to obtain the spatial information. It has been reported that the tree structure such as tree height and crown volume can be estimated from the LiDAR point cloud. To utilize the LiDAR data widely, automatic detection in the 3D point cloud is required. Trees in 3D point cloud can be detected using, for example, watershed and valleyfollowing method. Additionally, high classification and object detection accuracy could be performed using deep learning-based technique. In this study, we combined image processing and 3D deep learning technique to automatically detect trees in 3D point cloud obtained from airborne LiDAR.

In recent years, the deterioration of bridges has become a big social problem in Japan. Therefore, vigorous research is being carried out to detect the deterioration of bridges by measuring vibrations. Among them, there is a method of measuring vibration using images. The purpose of this paper is to enable high-accuracy measurement without contacts and artificial markers from a long distance. As a method, we took a video of the vibration with a stereo camera and measured it using frame-divided images. In addition, the resolution of the image was increased by using the super-resolution technolog y by deep learning, the accuracy of the corresponding point search by the AKAZE feature was improved, and the accuracy in long-distance measurement was improved.

Ultrasonic nondestructive visible testing using wavefield imaging has been applied to nondestructive testing by acquiring full-waveform data over a region of interest. Using the wavefield visualization, we can detect surface and sub-surface flaws in the target object. In this case, it would be useful that flaws can be automatically classified. In this study, a deep learning method based on the convolutional neural network (CNN) is applied to the classification of the flaw from the wavefield data. It is not easy to acquire a lot of the wavefield data in advance in actual fields. Therefore, we make numerical wavefield datasets as a substitute for the measured data using simulation by the elastodynamic finite integration technique. After building the CNN model, we used test data to verify the performance of the model. The CNN model using both simulation and measured dataset showed a high accuracy rate for the flaw classification.

Application of the machine learning methods is attempted to estimate the adhered salt on concrete bridges. The teaching data are obtained through a series of numerical simulations consisting of the three-dimensional fluid analysis around bridge and the random walk analysis. As the machine learning approach, the random forest regression and other four algorithms are considered. Based on the estimation accuracy, the performance of each method is evaluated. Moreover, the capability of the random forest regression is examined in detail, in the context of its application to the present purpose.

Recently, it is very important to reduce life cycle costs of bridges. Particularly, costs for prevention of the corrosion and rust comprise a large percentage of maintenance costs and the use of weathering steels for bridges has increased considering its cost-effectiveness. For example, the use of weathering steels is over 15% of new bridges in Japan. For maintaining of weathering steel bridges, estimation of the corrosion state is essential. The visual external inspection of the rust by skilled engineers is a common way for estimation of the corrosion. However, it is well known that there are some errors among the estimated results by different surveyors; hence, a simple and accurate evaluation method of the corrosion state is desired and deep learning has been paid attention. In this study, the convolutional neural network based on camera images about the rust visual inspection of weathering steel bridges is employed for quantitative estimation of the corrosion state. The results demonstrate that the proposed method can provide acceptable estimation of the rust conditions for practical application.

This paper aims to assess the steel deterioration through the images, which is the obtained data by the advanced technologies the corrosion tests and image classification by Convolutional Neural Network (CNN) analysis were conducted. From the corrosion tests results, there was a relationship between the exterior appearance and the gained weight of the corroded steel specimens. Therefore, the classification was conducted based on the gained weight of the specimens. CNN analysis was conducted by VGG19 classifier. As the result, to classify the images obtained from other environments by one classifier, the accuracy was low. Therefore, the classification based on images is concluded as difficult for steel corrosion deterioration. Thus, in order to adopt to real bridges, the images used to construct the classifier should be obtained from each bridges for suiting to each environment.

It ia an urgent issue to introduce new technology for demanding efficient and low budget with saving lavor since the bridge inspection is increasing every year. In recent years, we have applied Convolutional Neural Network (CNN), which is one of the machine learning that has focused an attention on its use in the field of civil engineering. CNN is considered to be one of the highly effective method of the support for bridge inspection. In this study, we developed a learning model that is a corrosion detector for steel girder bridges using CNN as machine learning. Our learning models trained using the photographs of the results of road bridge inspections conducted by Fukushima Prefecture. The corrosion detector derived from our learning models as was validated by using test data from the photographs of the ground survey at the road bridge in service of Inawashiro, Fukushima.

The infrared thermography method takes a thermal image of the concrete surface and detects internal defects. However, the thermal image diagnosis results may vary from person to person depending on the experience of the investigator, and improving the reliability of the inspection results is an issue.

In this study, we report on the technique for automatically identifying by statistical machine learning damage candidates extracted from thermal images. And report on the study results of automatic damage identification technology for infrared thermography method for bridge superstructure concrete and utilization status of thermal imaging diagnostic cloud service.

Drain of rain water plays a very important role for the prevention of corrosion for either of steel and concrete structures. When the drainage system does not work sufficiently, for example in case of abutment of a bridge, dust and/or sand absorbs the water resulting in not only aesthetic point but also long life of the structure. This study aims at developing water leakage detection system with artificial intelligence (AI). In particular, this study focuses on the elbow of the drain pipes, because of the joints where is a deterioration prone area. In addition, elbow part of pipes is subjected to momentum of the water. 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. Among algorithms of the machine learning, random forest was applied. The judgment was just whether water leakage occurs or not. However, the result of F value of 0.92 was obtained that is a significance value of the test.

Posting to social networking services during a disaster includes information that is useful for rescue and evacuation, but it is still underutilized in information gathering. In this study, we constructed a deep learning model to determine whether the posts containing keywords related to the disaster are valid or not. In addition, we visualized the words that the model focuses on. The mapping was made possible by extracting the location information from the post. It is shown that the built Deep Learning model can classify the submissions with high accuracy. The mapping was shown that the location information was generally extracted correctly. This suggests its effectiveness in classifying posts during disasters.