| Author: | Chen, Qianyi |
| Title: | Physics-informed machine learning for the monitoring of engineering structures |
| Advisors: | Cao, Jiannong (COMP) Zhu, Songye (CEE) Xia, Yong (CEE) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Deep learning (Machine learning) Structural health monitoring Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Computing |
| Pages: | xviii, 157 pages : color illustrations |
| Language: | English |
| Abstract: | Recent advancements in machine learning (ML) offer researchers alternative methods to interpret the physical world solely using data. These techniques have frequently surpassed traditional model-based methodologies across diverse fields, from physics and chemistry to biology. As a result, there’s a growing preference for ML in tackling scientific enigmas, especially those with undiscovered physical processes. However, existing ML methods still suffer from several limitations. Their ability to generalize is often restricted by the unlabeled and unbalanced training samples, particularly given the inherent challenges in collecting comprehensive data for every possible real-world situation. Furthermore, these methods can, at times, yield physically implausible predictions, violating the governing laws of physical systems. In light of these challenges, this thesis delves into physics-informed machine learning (PIML). The focal application is the monitoring of engineering structures. We introduce a framework that highlights various strategies for integrating ML models with diverse facets of physical knowledge, specifically in the context of structural health monitoring (SHM). This encompasses inherent parameters, architecture, and underlying physical equations of the monitored system. We first integrated physical parameters into a data-driven compression technique, ensuring vital physical quantities were retained during the lossy data compression. Specifically, we proposed physics-enhanced principal component analysis (PPCA) for compressing the vibration data of engineering structures in the context of structural health monitoring. By harnessing the intrinsic physical parameters from a structure, the compression process was expertly directed to conserve mode shapes, which are crucial vibration properties of structures. The indispensability of physical parameters in maintaining mode shapes is formally analyzed and theoretically proved. Empirical tests on both simulated and real-world structures revealed that PPCA can improve the accuracy by up to 56% compared with the alternative baseline. We further integrated prior architectural knowledge of physical systems into data-driven imputation methods. It leads to the creation of the multi-stage graph convolutional network with spatial attention (MSA-GCN) for recovery of missing values in multivariate time series (MTS). MSA-GCN uniquely aligns with the multi-stage graph architecture of heterogeneous monitoring data. In the first stage, we decompose heterogeneous MTS into several clusters with homogenous readings from similar sensor types and learn intra-cluster correlations. Subsequently, a GCN with spatial attention is employed, designed to discern dynamic inter-cluster correlations—representing the second stage of MSA-GCN. The final stage decodes features from preceding stages using stacked convolutional neural networks. We jointly train these three-stage models to predict the missing data in MTS. The joint use of multi-stage design and the spatial attention mechanism empowers MSA-GCN to adeptly understand the heterogeneous and dynamic correlations in MTS, yielding enhanced imputation results. We demonstrated the efficacy of MSA-GCN using real-world data from a large-span bridge in Hong Kong and showcase the superiority of MSA-GCM against the baselines in achieving the lowest imputation error. We explored the implementation of PIML in learning the dynamics of physical systems. We introduced an innovative data-driven simulator called physics-informed edge recurrent simulator (Piers) for learning the dynamics of continuous deformable bodies (CDBs). We first formulate the CDB simulation as a sequence-to-sequence input-output relationship modeling problem. Then we incorporate a recurrent neural network (RNN) to learn edge updates within short timesteps, during which stiffness changes are negligible. For enhanced accuracy, the RNN module’s hidden states are initialized with inherent physical knowledge. Additionally, Piers integrates a physics-informed loss function, targeting edge output based on the actual physical properties of interactions. Extensive experimental results confirm Piers’ efficacy in predicting the dynamics of various CDBs—be it elastic, plastic, or elastoplastic—with significantly improved accuracy over other baselines. Specifically, in simulations across four typical CDBs, Piers reduces prediction errors by an impressive range of 78% to 99%. This thesis pioneers innovative data-driven methods for monitoring real-world engineering structures. Constrained by physical prior knowledge, ML models are trained within a refined search space, leading to efficient model training. Given the universality of physical knowledge across scenarios, the generalization capabilities of PIML models are enhanced, substantially reducing the likelihood of making physically implausible predictions. Comprehensive simulation data, real-world measurements, and experiments underscore the exceptional performance of the proposed PIML methods. Looking ahead, PIML holds promise for broader applications in monitoring other physical systems where both measurement data and established governing principles are at hand. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- 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.
- 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.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/12809