This research aimed at applying revolutionary and transferable technologies, including the
advanced sensing and 3D printing technologies, and artificial intelligence, to some of the
problems in civil engineering. Three different applications, i.e., the determination of
consolidation parameters, microstructural characterizations of soil samples and health
monitoring of the civil infrastructures, were studied in this thesis.
In the first part, the U-oedometer, a novel modified oedometer cell equipped with tailor-made
needle probes, is developed to easily and accurately measure the excess pore water
pressure (Δ?) during 1-D consolidation tests, and to determine the coefficient of consolidation
(?
?). The 3D printing technique is applied to make simple yet robust modifications to the...[
Read more ]
This research aimed at applying revolutionary and transferable technologies, including the
advanced sensing and 3D printing technologies, and artificial intelligence, to some of the
problems in civil engineering. Three different applications, i.e., the determination of
consolidation parameters, microstructural characterizations of soil samples and health
monitoring of the civil infrastructures, were studied in this thesis.
In the first part, the U-oedometer, a novel modified oedometer cell equipped with tailor-made
needle probes, is developed to easily and accurately measure the excess pore water
pressure (Δ?) during 1-D consolidation tests, and to determine the coefficient of consolidation
(?
?). The 3D printing technique is applied to make simple yet robust modifications to the
conventional oedometer cell for facilitating the installation of the needle probes. The tailor-made
needle probes are designed in such a way that the volumetric compliance is lowered to
avoid measurement bias. In addition, a power-regulating sensor board is used to stabilize the
voltage output for long-term measurements so as to improve the accuracy of the Δ?
measurements. Subsequently, the Δ? -based method is proposed to determine ?
?, through
minimizing the difference between the measured and theoretical values of Δ? . Such an
approach avoids the intervention of human judgement and therefore minimizes the degree of subjectivity. The experimental results demonstrate that the measured Δ? matches the
theoretical values of the Terzaghi 1-D consolidation theory, showing that the estimated ?
? is
sufficiently reliable to reflect the whole consolidation process. In addition to the determination
of ?
?, the U-oedometer allows additional measurements of other soil properties during
consolidation, including the coefficient of permeability (?) and the coefficient of earth pressure
at rest (?
0). It is observed that k decreases with the reduction of void volume, due to the increase in effective vertical stress (?
?′). Further, the secondary compression seems to be a continuation
of the primary consolidation, where the soil sample continues to deform at a relatively slower
rate, associated with the slight decrease in ?. A constant value of ?
0 is observed at any value
of ?
?′ in the loading path, while during secondary compression, ?
0 slightly increases with time.
The second part reports the use of the Deep Learning-based technique to rapidly,
automatically and objectively characterize the microstructure of clay samples. The U-Net
model was applied to perform semantic segmentation for identifying individual kaolinite
particles, based on the Scanning Electron Microscopic (SEM) images taken from clay samples
subjected to 1-D consolidation. To facilitate supervised learning, the elongated particles with
known orientation were manually annotated first; additionally, data augmentation was
extensively applied to increase the size of the training dataset and 5-fold cross-validation was
used to ensure satisfactory generalization of the Deep Learning models on the unseen dataset.
Dice loss and weighted cross-entropy were chosen as the loss functions to tackle the issue of
imbalanced classification class. The customized weight maps incorporated in the weight cross-entropy
were found effective in forcing the Deep Learning models to learn how to recognize
the particle boundaries. With the trained Deep Learning models, the targeted kaolinite particles
(>~12000 particles) were identified from the SEM images within ~20 mins and the particle
directional distribution was quantified using the fabric tensor. The characterization results
reveal that the kaolinite particles exhibit a tendency to gradually align along the horizontal
plane as imposed by the applied vertical stress.
In the first section of the third part, a convolutional autoencoder was used to implement
an unsupervised anomaly detection of the damage of concrete structures from images, so as to
facilitate the condition assessment of the civil infrastructures. Upon the completion of training,
this novel approach renders poor reconstruction of the defects of concrete structure, in turn,
aids detecting the location of defects. Since the convolutional autoencoder was trained to
reconstruct images, no label was required, thereby saving enormous time for annotating labels.
Comparison was carried out with the segmentation results produced by other automatic classical methods. The analysis reveals that results made by the anomaly map generated by the
proposed method outperform other segmentation methods, in terms of precision, recall, F
1
measure and F
2 measure, without severe under- and over-segmentation. Further, instead of
merely being a binary map, each pixel of the anomaly map is represented by the anomaly score,
which acts as risk indicator to alert inspectors, wherever the damage of concrete structures is
detected.
In the second section of the third part, deep learning was explored to facilitate the health
monitoring of the civil infrastructures, in terms of augmenting the degree of automation and
minimizing the needs of labor-intensive processing works of the defect detection and
classification. A classifier was first designed and trained from the scratch, for categorizing
images to the classes of no defects, cracking and spalling. Upon the completion of the model
training in a 5-fold cross-validation fashion, the built classifiers achieve satisfactory
performance in categorizing images into the classes of no defects, cracking and spalling.
Following that, an artificial intelligence-empowered monitoring pipeline was constructed. An
unsupervised screening and detection of defects, made by the aforementioned convolutional
autoencoder, was carried out in advance of the defect classification. This anomaly detector
substantially reduces the search space of defects, in turn, allows more computational resources
and time can be allocated on analyzing the defects of the concrete structures, which are the
main focus of the task. Comprehensive analyses show that the proposed AI-empowered
monitoring pipeline is capable and adaptable in detecting and classifying defects subjected to
a wide variety of environmental conditions, including lighting condition, camera distance and capturing angle.
Post a Comment