Operational structural damage identification using de-noised modal feature in machine learning / Chen Shilei
Structural damage can severely affect the safety and functionality of the structure and lead to economic loss. Vibration-based structural damage detection has raised continuous interest over the decades, as a non-destructive way to provide warnings and predict certain faults at early stages. Comp...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2021
|
| Subjects: | |
| Online Access: | http://studentsrepo.um.edu.my/13198/ http://studentsrepo.um.edu.my/13198/1/Chen_Shilei.jpg http://studentsrepo.um.edu.my/13198/8/shilei.pdf |
| Summary: | Structural damage can severely affect the safety and functionality of the structure and
lead to economic loss. Vibration-based structural damage detection has raised continuous
interest over the decades, as a non-destructive way to provide warnings and predict certain
faults at early stages. Compared with conventional modal parameters such as the natural
frequency and mode shape, the upstream modal data, namely the frequency response
function (FRF), can be a better alternative, because it is rich in modal information and
can be easily obtained. However, the FRF is usually measured through experimental
modal analysis (EMA) when the test object is in shut-down mode, which is not practical
for real-time application in the working environment. This limitation can be overcome by
a novel technique named impact-synchronous modal analysis (ISMA) performed under
the operational condition. Machine learning is also a focus in this work, which was
employed to process and classify FRF data in terms of damage. By integrating ISMA,
both supervised and unsupervised machine learning algorithms were investigated to
develop real-time damage identification schemes. Specifically, the back-propagation (BP)
network was employed in the supervised learning method, and the FRF changes in a
selected frequency interval at several measurement points were used as the input of the
network. The unsupervised learning method was developed by combining principal
component analysis (PCA), waveform chain code (WCC) analysis and hierarchical
cluster analysis. WCC analysis was carried out on the PCA-reduced FRF to extract
damage-sensitive PCA-WCC features. The unsupervised hierarchical cluster analysis was
then conducted on these features. The proposed schemes were tested on a rectangular
Perspex plate. The results show that the similarity between the FRF obtained by ISMA
and EMA exceeds 0.993, proving that the de-noising method of ISMA provides static
comparable FRF data during the in-service condition. For the supervised learning method,
the trained BP network can successfully identify the scenarios of high and moderate
damage with an overall accuracy of 100% when all five measurement points are used.
With the input features optimized by mode shape assessment, 100% accuracy can also be
achieved with only two measurement points. For the unsupervised learning method, the
hierarchical cluster analysis can correctly cluster the samples in terms of their damage
states. In terms of damage severity and location identification, the proposed scheme is
sensitive to detect damage severity with damage index as low as 0.05. In addition, the
combination of PCA-reduced FRF and mode shapes shows a positive correlation between
the magnitude of the resonant peak and the displacement of the impact point in identifying
the damage location of the plate. In conclusion, the supervised learning method using
FRF change is convenient and effective in identifying the damage state of the plate, and
can be optimized through mode shape assessment. Meanwhile, the unsupervised learning
method using PCA-WCC features is good at detecting unknown damage, and is sensitive
to low-severity damage. With the help of PCA-reduced FRF, it is also feasible to estimate
the severity and locate the damage of the test plate. |
|---|