2022 Fiscal Year Final Research Report

High Accuracy Evaluation of Concrete Delamination Risk by Deep Learning of Hammering Sounds

Research Project

PDF

Project/Area Number	20H02234
Research Category	Grant-in-Aid for Scientific Research (B)
Allocation Type	Single-year Grants
Section	一般
Review Section	Basic Section 22020:Structure engineering and earthquake engineering-related
Research Institution	Kyushu University
Principal Investigator	Sonoda Yoshimi 九州大学, 工学研究院, 教授 (40304737)
Co-Investigator(Kenkyū-buntansha)	玉井宏樹九州大学, 工学研究院, 助教 (20509632) 別府万寿博防衛大学校(総合教育学群、人文社会科学群、応用科学群、電気情報学群及びシステム工学群), システム工学群, 教授 (90532797)
Project Period (FY)	2020-04-01 – 2023-03-31
Keywords	打音検査 / コンクリート構造物 / 劣化損傷（浮き・剥離） / 深層学習 / 畳み込みオートエンコーダー / 畳み込みニューラルネットワーク
Outline of Final Research Achievements	In this study, a hammering sound test is performed on concrete specimens with artificial defects and recorded. The followings are found.1)The presence or absence of defects can be ascertained with nearly 100% accuracy for both Convolutional autoencoder and Convolutional neural network that learned the timefrequency analysis result of the tapping sound data, 2)It is possible to detect minor defects that were difficult to detect with indexes such as the Amplitude ratio. 3)There is a correlation between the reconstruction loss value of CAE and the degree of damage, and it is possible to predict the degree of damage by CAE, however deeper the defect position, the more blurry it becomes, and accuracy deterioration is inevitable. 4)CNN is superior to CAE in diagnosing the degree of damage (range and depth), but since the quality (bias) and range of the training data have a great influence on theie accuracy, it is necessary to examine the appropriateness of the training data in advance.
Free Research Field	Structural Engineering
Academic Significance and Societal Importance of the Research Achievements	本研究ではCAEとCNNの２種類の手法で打音の時間周波数スペクトラムを用いた深層学習を行い，両者の診断精度について比較検討を行った。その結果を要約すると以下のようになる。 1)CAE, CNNどちらの手法も高精度でコンクリート内部の浮き・剥離の有無を把握できることが確認できた。2)CAEは欠陥の有無だけでなく損傷の程度も評価可能ではあるが，欠陥が深くなるにつれて精度が低下するだけでなく，周囲騒音の影響も受け易いことが認められた。 3)CNNは適切な学習データを収集・学習させることで，CAEより精度が高い損傷程度の評価が可能であるだけでなく，周囲騒音の影響も受けにくいことが認められた。