THESIS
2020
Abstract
A key step of real-time structural health monitoring in plates is the fast damage
localization in a large area for further inspection. With the guided wave signals sensed at several
locations, the damages can be localized directly by time-of-flight based and damage index-based
imaging algorithms. However, the former requires decomposition of complicated signals,
and the latter needs a dense array of transducers that restrict their application scope.
Alternatively, model-based or machine learning-based methods localize the damage of an
unknown signal referring to a dictionary of signals with known damaged states, a.k.a. training
data. These algorithms can localize damage without the interpretation of signals and a dense
sensor array as long as a high-quality and sufficient dictio...[
Read more ]
A key step of real-time structural health monitoring in plates is the fast damage
localization in a large area for further inspection. With the guided wave signals sensed at several
locations, the damages can be localized directly by time-of-flight based and damage index-based
imaging algorithms. However, the former requires decomposition of complicated signals,
and the latter needs a dense array of transducers that restrict their application scope.
Alternatively, model-based or machine learning-based methods localize the damage of an
unknown signal referring to a dictionary of signals with known damaged states, a.k.a. training
data. These algorithms can localize damage without the interpretation of signals and a dense
sensor array as long as a high-quality and sufficient dictionary is given. However, in practice,
limited training data are normally available generated by structures or models that are not
entirely identical to the monitored structure. This brings the challenges of transferability,
robustness, and interpolation ability to these methods while most of them in literature weren’t
demonstrated in these aspects.
To overcome these challenges and develop a solution without the needs of interpretation
and dense array, this work integrates feature extraction techniques of guided waves with
machine learning algorithms. The feature extraction is under the principle of reserving temporal
information that is crucial to damage localization while reducing the dimensionality and
removing superfluous information.
A modified model-based method with dimensionality reduction and interpolation is
developed at first to improve its performance under limited data. Then, a 1-D convolutional
neural network (CNN) is proposed to overcome all the aforementioned challenges.
In three experiments, these methods are demonstrated to overcome the challenges and
obtain accurate localization using complicated signals from a sparse sensor array. Their
superiority to the literature and the advantage of extracting features using 1-D CNN are also
demonstrated.
Post a Comment