| Project/Area Number |
21K17829
|
|---|
| Research Category |
Grant-in-Aid for Early-Career Scientists
|
| Allocation Type | Multi-year Fund |
|---|
| Review Section |
Basic Section 61050:Intelligent robotics-related
|
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2021-04-01 – 2024-03-31
|
| Project Status |
Completed (Fiscal Year 2023)
|
| Budget Amount *help |
¥4,550,000 (Direct Cost: ¥3,500,000、Indirect Cost: ¥1,050,000)
Fiscal Year 2022: ¥1,820,000 (Direct Cost: ¥1,400,000、Indirect Cost: ¥420,000)
Fiscal Year 2021: ¥2,730,000 (Direct Cost: ¥2,100,000、Indirect Cost: ¥630,000)
|
|---|
| Keywords | Inspection / Domain adaptation / Clustering / Weak supervision / Machine learning / Non-destructive testing / ドメイン適応 / インフラ点検 / コンクリート構造物 / 非破壊検査 / 機械学習 |
|---|
| Outline of Research at the Start |
The aging of concrete structures is an issue worldwide and the automation of inspection methods such as the acoustic hammering test is highly demanded. To achieve this, a model is learned using labeled data. However, labeled data must correspond to the inspected structure, which is highly impractical.
The aim of this research is to realize a model for defect detection in concrete structures that can learn from a reference structure on which labeled data is available and generalize to the inspection target structure, for which labeled data is not available.
|
| Outline of Final Research Achievements |
Concrete structures make up a large portion of buildings in modern societies. This is especially true for social infrastructures (tunnels, bridges, highways,...). Their regular inspection is critical to ensure the safety of the public. The hammering test is an inspection method based on acoustic means that is currently popular but is man-power heavy. Between the growing number of concrete structures in need of testing and the man-power shortage, the automation of the hammering test is highly desirable.
In recent years, machine learning models have made great strides, showing great performance on recognition tasks. However, most require large amounts of labeled training data to do so. This is not feasible for niche applications such as the hammering test. Therefore, this research has focused on ways to allow models to perform well independently of the available training data.
|
| Academic Significance and Societal Importance of the Research Achievements |
Several avenues were considered. Multi-modal domain adaption, domain expansion and domain independent data extraction. Those work on different levels and great results were obtained for each. In the future, their integration into a single framework for a working system could be considered.
|
