In recent years, flood damage caused by heavy rain has become frequent, thus necessitating rapid understanding of levee overflow to respond to disasters promptly and disseminate information to the local area. In this study, we constructed three models that can detect overflow events using large-vision-language models and then clarified the detection accuracy of the developed model using actual data (image data from heavy rainfall caused by Typhoon No. 19, 2019). By performing fine-tuning based on OpenAI’s GPT-4o, we demonstrated the possibility of accurately distinguishing between overflow events and normal times.

Recently, nondestructive evaluation techniques using deep learning have attracted attention, and structural damage inference methods employing convolutional neural networks (CNNs) have been proposed. However, classification accuracy tends to degrade when the distribution of test data differs from that of the training data. In this study, simulated damage conditions were reproduced by drilling holes in wooden specimens, and a CNN model was constructed to classify the damage into three levels. To develop a generalizable model capable of accurately classifying not only the same type of wood but also different types, a domain adaptation method using a Domain-Adversarial Neural Network (DANN) was introduced and its effectiveness was validated. The results showed that semi-supervised learning, which incorporates a portion of labeled target domain data into training, significantly improved classification accuracy.

This study summarizes the results of the Image Analysis WG1, which was held a total of nine times over a period of approximately two years from March 2023 within the AI and Data Science Practical Research Subcommittee of JSCE. The study mainly looked at image classification, which is relatively easier to handle than object detection, and the use of apps that can be used with no-code or low-code, as well as the impact of cropping the training data on the classification of damaged images. As a result, it was found that no-code apps are very useful depending on the subject, and that the effect of cropping images is not that great.

Image processing has been applied to various fields of engineering, including medicine, industry, man- ufacturing, transportation and autonomous driving, as well as crime prevention and surveillance. In the 2010s, advances in deep learning significantly enhanced the accuracy and versatility of image processing, leading to a wider range of practical applications. In the first half of this paper, we highlight some of the most widely used image processing techniques, including those based on deep learning, and describe their development. While these image processing techniques are powerful tools, they are not all-purpose solu- tions; users must tailor their applications according to specific objectives. Therefore, in the latter half of this paper, with the aim of contributing to the future application of image processing in the civil engineering field, we introduce several case studies where the authors applied image processing techniques to real- world tasks in civil engineering, particularly in structural engineering.

This paper reports on the activities of the Physical Model + AI Working Group established within the AI and Data Science Practical Research Subcommittee of the Structural Engineering Committee in the Japan Society of Civil Engineers. Aiming to integrate AI technology and physics, this WG conducted research mainly on a method called Physics-Informed Neural Networks (PINNs), which has attracted particular attention in recent years. Specifically, the working group conducted a literature survey on PINNs, examined the learning convergence of PINNs, applied PINNs to the analysis of marine tsunami propagation, performed inverse analysis of wave propagation problems using PINNs, and applied neural networks as surrogate models to mechanical problems. This paper provides an overview of the results of these studies.

This report has been prepared as an activity report of the Working Group within the Structural Engineering Committee’s AI & Data Science Practice Research Subcommittee, which examined the use of language models. With the emergence of large-scale models such as LLMs and Omni models capable of handling images and data analysis, the investigation focused on the potential outcomes and challenges of applying these cutting-edge technologies to research and technological development in the civil engineering field. The examination was conducted by organizing papers submitted by members during the project period, and recommendations were provided based on approaches from other fields to address the identified challenges.

Working Group 6 of the Subcommittee on AI Applications in Structural Engineering conducted activities over a period of three years focused on the utilization of 3D models for infrastructure structures. These activities primarily involved investigating and researching the current state of 3D reconstruction of existing structures and the application of artificial intelligence technologies, as well as sharing information and exchanging opinions.

This paper summarizes and reports the findings obtained from the working group activities regarding the effectiveness and future potential of 3D models for infrastructure structures.

This study constructed an integrated dataset centered on the nationwide road facility inspection database (xROAD), combining structural information, meteorological data, and maintenance records, and applied AI methods to analyze bridge deterioration factors. For steel bridges, the damage score per unit area (Saa) was introduced, and explainable machine learning models incorporating principal component analysis (PCA) and SHAP were employed to clarify the key features influencing corrosion progression. For concrete bridges, inspection documents, specifically the “Comprehensive Inspection Results” section, were analyzed using morphological analysis and tf-idf, revealing the relationship between textual information and changes in bridge condition.

This study focuses on time-series vibration data of bridges and applies AI-based analysis methods for assessing structural conditions and traffic environments. Acceleration response data were utilized to esti- mate girder deflection and classify traffic types, highlighting the potential for reliable data analysis and practical implementation. A machine learning model was constructed to correct errors arising from double integration in displacement estimation, improving accuracy. In addition, a method for detecting vehicle entry and exit times using longitudinal acceleration and an extended neural network approach for traffic classification were proposed. The results suggest that these approaches can enhance traffic census accuracy and promote more effective utilization of existing infrastructure.

In recent years, various data platforms for infrastructure facilities have been developed and put into operation, such as the road data platform developed by the Ministry of Land, Infrastructure, Transport and Tourism and PLATEAU, which provides open access to 3D city models. The “Data Integration/Management + AI Working Group” of the AI and Data Science Practice Research Subcommittee organized use cases and challenges of point cloud data and AI-based solutions to these challenges as essential technologies for realizing infrastructure facility data management by combining data platforms and AI. In addition, the group conducted performance evaluations of foundation models such as VLMs and LLMs using infrastructure facility data.

In recent years, the deterioration of sewer infrastructure has been progressing, increasing the importance of maintenance and management. At the same time, the aging of skilled engineers and a growing shortage of personnel have made it increasingly difficult to secure experienced human resources. As a result, the number of young and newly assigned engineers is increasing, but the knowledge required for sewer maintenance is dispersed across a vast amount of documentation, making it difficult to quickly obtain necessary information. To address this issue, this study aims to develop a chatbot that can efficiently provide expert knowledge related to sewer maintenance and management. Leveraging large language models (LLMs), we examined three approaches: a retrieval-augmented generation (RAG) method, a fine- tuning method, and a hybrid method combining both. After preparing a structured dataset, we conducted functionality checks and performance evaluations through simulated dialogues. The results indicated that the hybrid approach combining RAG and fine-tuning achieved the highest response accuracy and stability, suggesting its potential usefulness in practical sewer maintenance operations.

The 2024 Noto Peninsula Earthquake and subsequent torrential rains in Oku-Noto caused extensive damage to social infrastructure across the region. Given the critical role of road networks in supporting local industries and tourism, rapid restoration is imperative. However, the peninsula’s unique geographical and infrastructural characteristics, coupled with the widespread nature of the damage, pose significant challenges to conventional inspection methods in terms of time, manpower, and safety. To address these issues, the Noto Reconstruction Office implemented a novel inspection approach integrating AI technologies with UAVs. Field investigations demonstrated that this method offers high diagnostic accuracy and substantial time savings, confirming its effectiveness for post-disaster infrastructure assessment.

Recent research has explored the automation of hammering-based nondestructive testing for reinforced concrete structures using machine learning techniques. However, many of these studies have not been directly validated against internal defects in actual structures.

In this study, hammering sound data were collected from large-scale specimens in which corrosion- induced cracks were reproduced through electrolytic degradation. Unsupervised anomaly detection using three types of autoencoders was applied, and the results were compared with expert judgments and core sampling.

The findings suggest that applying continuous wavelet transform to hammering sound data enables the identification of critical deep defects that may be overlooked by experienced inspectors. Furthermore, the proposed method achieved highly accurate anomaly detection without extensive hyperparameter tuning or expert labeling, indicating its potential to improve the efficiency of nondestructive inspection.