請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/69394
標題: | 應用機器學習與影像量測於結構健康監測:模態參數、損傷診斷及裂縫辨識 Applications of Machine Learning and Image Processing to Structural Health Monitoring: Modal Parameters, Damage Diagnosis, and Crack Detection |
作者: | Jung-Wen Yu 尤俊文 |
指導教授: | 張家銘(Chia-Ming Chang) |
關鍵字: | 結構健康監測,機器學習,電腦視覺,影像處理,影像量測,自然頻率估計,模態振型萃取,遞移函數相位損傷識別,裂縫自動辨識, Structural Health Monitoring,Machine Learning,Computer Vision,Image Processing,Image Measurement,Natural Frequency Estimation,Mode Shape Extraction,Transfer Function Phase-Based Damage Diagnosis,Crack Recognition, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 近年來,為了有效延長建築結構使用的生命週期,及保護其免於受到災害再次破壞,結構健康監測成為結構物永續經營最關鍵的部分。由於電腦技術的普及與進步,透過即時與自動化之方式,將結構物之健康資訊傳遞給工程師,以利後續之補強工程,是良好的結構健康監測行為。然而,在傳統的結構健康監測方式中,常利用有線感測器,傳遞電子訊號至主機,再由主機從感測器獲得資訊,診斷結構當前之健康狀況。另外,為了強化結構健康監測的手段,並有效應用目前影像處理計算優勢,本研究藉由相機獲取結構動態影像,結合影像量測和電腦視覺的方式處理動態影像,估計被拍攝結構的位移變化,並獲取自然振動頻率與模態振型等模態參數,進而了解結構當前的健康狀態。 人工智慧因目前電腦硬體效能有所提升,可即時、迅速處理大量影像數據,有機會改善過去人工巡檢的缺點,不僅可以節省人工判斷的時效性,亦可改善誤判的情形。若能以人工智慧技術對結構反應,即時判斷結構損傷的程度,估計結構殘餘性能,則可大幅改善救災的規劃,並保護使用者生命、財產的安全。因此,本研究基於傳統機器學習流程,應用人工智慧深度學習結構的反應特性,得到判斷損傷發生、位置,並瞭解結構當前的性能。由於建築結構會因為時間老化問題而導致剛度以及強度下降,甚至造成建築物損壞,而產生一連串的公共安全問題。因此若能建立一套有效且簡易的深度學習模型,將結構受損前後的遞移函數之相位差,以傳統的神經網路中學習,當結構物損傷後,能夠在第一時間檢測出損傷的位置以及程度,便能夠更快速的進行補強和救災。 另外,本研究同時也利用深度學習判斷混凝土構件表面之裂縫性質,以非破壞性檢測裂縫之方式,提供裂縫相關資訊,加速結構工程師判斷混凝土構件之健康情況。基於人工智慧中深度學習以及遷移學習的方式,再藉由學習訓練資料中每張影像特徵,訓練出屬於本研究之裂縫辨識模型。訓練完畢後,透過其架構能夠提取特徵之優勢,進而自動化地判斷影像中混凝土表面是否有裂縫存在,並框選出裂縫之位置,並同時在此框選區域中以遮罩的方式輸出此裂縫,隨後將深度學習所遮罩出的裂縫位置,進行影像後處理,經過一系列電腦視覺之方法,將混凝土表面上的裂縫萃取出,將裂縫完整的標記回原始影像上。 Structural health monitoring (SHM) has gained more attention in the field of civil engineering for the purpose of structural damage prevention and long-term structural maintenance. With the advanced computer technologies, structural soundness can be accessed through derived monitoring indicators. These indicators can assist engineers for engineering judgements such as retrofitting performance and damage levels. In the past, SHM applications relied on wired sensors to access structural information and then to record the readings by data acquisition systems (DAQ). However, installing wired or surface-contact sensors may result in substantial challenges if complex, large dimension structures are considered to be instrumented. Moreover, the installation process is tedious and time consuming. Alternatively, still or motion cameras can be employed as non-contact sensors which still yield useful information for structural health monitoring. In this research, dynamic responses are recorded through a single camera device. Modal information of structures (or structural components) can be obtained through regional intensity variations of images. Also, the displacements of certain region of interest (ROI) can be estimated by super-resolution imaging techniques. These displacements responses are then utilized for modal parameter extraction, e.g., mode shapes, natural frequencies, and damping, to access the current structural conditions. With the advance in computational development, complex and heavy-loaded computations (e.g., image processing) have become more effective than before. One of the examples is artificial intelligence techniques which are now more suitable for real-world applications. With the aid of machine learning, structural integrity can be investigated through structural responses and represented as a model trained by a self-learning architecture. The model is capable of either categorizing the structural state (i.e., healthy or damaged) or estimating the structural residual performance. In this research, a conventional machine learning process is applied for structural damage detection. Transfer function phases are visualized and selected as the training input. Meanwhile, the damage location, damage level, residual performance are chosen as the output. This research can expedite the maintenance process after earthquake events and track long-term structural conditions. Additionally, machine learning can also be utilized as nondestructive evaluation (NDE) of crack detection on concrete surfaces. The conventional crack detection requires engineers to carefully examine each location of a structure; however, using machine learning techniques render an opportunity to exploit continuous image data (e.g., frames from a video camera) for crack detection. As the convolution neural network has developed, a deep learning model can be trained and established from acquired images which have the same features (e.g., concrete surface cracks). Then, the trained model can detect and highlight cracks in an image. Therefore, this research focuses on using Mask Region-based Convolutional Neural Network (Mask R-CNN) which generates a bounding box on detected crack region with object segmentation. Moreover, by combining with computer vison techniques, the extracted crack can be mapped on the original image, and the crack properties can be identified. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/69394 |
DOI: | 10.6342/NTU202003970 |
全文授權: | 有償授權 |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-1808202014311600.pdf 目前未授權公開取用 | 5.57 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。