Author: Chen, Sixin
Title: Generalizable deep learning for structural health monitoring : from graph formulation to domain adaptation
Advisors: Ni, Yi-qing (CEE)
Zhou, Lu (CEE)
Degree: Ph.D.
Year: 2022
Subject: Deep learning (Machine learning)
Medicine -- Data processing
Hong Kong Polytechnic University -- Dissertations
Department: Department of Civil and Environmental Engineering
Pages: xxxiv, 321 pages : color illustrations
Language: English
Abstract: In recent years, the rise of deep learning (DL) has boosted studies on DL-based structural health monitoring (SHM), which opens up a new direction for this area. DL-based SHM does not require a detailed physical description of the structural system, unlike model-based methods or hand-crafted features sensitive to damages, unlike other data-driven methods. Despite of these advantages, a DL model requires great efforts for data collection and model training, so it is desired to be applicable to new cases even under different conditions. However, the condition differences, including the shift of sensor arrangement, the lack of training samples and the change of environment, hinder the applicability of DL models. To overcome these issues, generalizable DL methods for SHM are developed based on graph neural networks (GNNs) and transfer learning (TL), demonstrated and validated through four case studies in mechanical and civil scenarios.
In the first case, graph convolutional networks (GCNs) are introduced to handle the shift of sensor arrangement. Specifically, a sensor network and multiple bolts, as well as their distances, are formulated as a graph, on which the proposed EMI-GCN can infer the loosening of multiple bolts by propagating the information of impedance variation. Thanks to the explicit encoding of distance effect and the automatic selection of frequency sub-range sensitive to loosening, the EMI-GCN performs well even when the sensor arrangement has shifted.
As an advancement of EMI-GCN, a new model is proposed in the second case, which does not require the locations of damages for graph formulation and further utilises convolutional neural networks for feature extraction. A sensor network and the potential damages are formulated as vertices of a heterogenous graph. The process of how baseline signal is altered by multiple potential damages is considered as the message passing between these vertices. The proposed UGW-GNN infers the severities and locations of multiple damages based on messages received by the receiver vertices. It is proved that UGW-GNN can conduct damage inference without relying on any specific sensor pair, making it flexible to the sensor arrangement.
In the third case, the lack of samples is handled by inductive TL. This strategy is utilised to induce the DL model for acoustic emission-based rail condition evaluation by knowledge learned from a sound event classification task. The number of parameters to learn can be reduced due to the task relevance, which improves the generalization capability of DL model compared with that learned from scratch. The contribution is to show that a more suitable source task can enhance time series data-based SHM practice and prove this rationality by introducing a metric called maximum mean discrepancy.
The change of environment is handled by transductive TL, i.e., domain adaptation (DA). The fourth case study presents a semi-supervised adversarial DA approach, which enables assessing the evolving wheel conditions for high-speed trains operating in different surrounding environments. In addition, a semi-supervision tactic is adopted to further align the conditional distribution of labelled data. This case study reveals the possibility of monitoring critical components of high-speed trains running in any rail line, given only the healthy data and a well-studied case in another rail line.
Rights: All rights reserved
Access: open access

Files in This Item:
File Description SizeFormat 
6183.pdfFor All Users41.8 MBAdobe PDFView/Open

Copyright Undertaking

As a bona fide Library user, I declare that:

  1. I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
  2. I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
  3. I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.

By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.

Show full item record

Please use this identifier to cite or link to this item: