Browsing by Author "Hadavandsiri, Zahra"

Now showing 1 - 1 of 1
  • Thumbnail Image
    ItemOpen Access
    Concrete Damage Inspection by Classification of Terrestrial Laser Scanner Point Clouds
    (2020-05-05) Hadavandsiri, Zahra; Lichti, Derek D.; Shahbazi, Mozhdeh M.; Dawson, Peter C.
    Concrete structures endure damage and deterioration when subjected to human activities and natural hazards. Early detection of damage on concrete structures is vital to counter the side effects deriving from damage and to allow timely maintenance procedures. This thesis presents a novel approach for high-precision detection of damage on concrete surfaces using terrestrial laser scanner point clouds (PCs). At first, an unsupervised approach is developed that utilizes a robust version of principal component analysis (PCA) classification in order to distinguish between structural damage and outliers present in the data. Numerical simulations are conducted to develop a systematic point-wise defect classifier that automatically diagnoses the location of surface damage on the investigated region. The developed method examined on two real datasets, demonstrate the validity of the proposed systematic framework for reliable detection of damage of any type which makes roughness as small as 1 cm or larger on the surface of concrete structures captured with any laser-scanning PC with a minimum spatial resolution of 5 mm point spacing. At second, a supervised approach is developed that employs the outcome of the primary unsupervised classifier in order to accurately annotate the training data without the need for manual labeling. One flume of an aqueduct dataset was used for training the system. This machine learning-based model relies on a support vector machine (SVM) algorithm to train a point-wise defect classifier for locating the concrete damage. This yields an average classification precision and F1-score of 97.33% showing the potential of using machine learning for concrete damage detection. The performance of the prediction model was evaluated on three real datasets. The prediction model can successfully mirror the high performance of the unsupervised method used in the training process. In addition, by exploiting a more extensive variety of geometric features and skipping the intensive computation of the robust PCA, it outperforms the unsupervised classifier in terms of model performance and computational efficiency, respectively. Consequently, the properly trained machine learning system provides reliable diagnosis of the health conditions of large concrete structures that are not computationally feasible to be inspected by the primary unsupervised classifier.