Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Data and informatics-informed digital twins for smart and practical bridge maintenance

Gao, Yan ORCID: 2023. Data and informatics-informed digital twins for smart and practical bridge maintenance. PhD Thesis, Cardiff University.
Item availability restricted.

[thumbnail of Final PhD Thesis_Yan]
PDF - Accepted Post-Print Version
Download (9MB) | Preview
[thumbnail of Cardiff University Electronic Publication Form] PDF (Cardiff University Electronic Publication Form) - Supplemental Material
Restricted to Repository staff only

Download (530kB)


Digital twinning has garnered significant interest for its potential to revolutionize bridge operation and maintenance (O&M) in the era of Industry 4.0. Although digital twin (DT) systems have achieved considerable development in academia and are getting increasingly popular in the industry, many gaps exist between academic research and widespread applications in route practices of bridge O&M, such as DT’s efficiency, resiliency, and intelligence, which are associated with different scenarios in a realistic bridge maintenance cycle. This research is conducted to solve the issues encountered in DT practical applications for bridge maintenance, and it can be structured according to five parts. Firstly, an efficient and resilient bidirectional DT framework is proposed for bridge O&M based on a comprehensive understanding of communication complexity by leveraging AI-informed edge computing, information hierarchy, and low-powered wide area network (LPWAN). The theoretical framework is idealised mathematically with state space representation and modelled using Petri-net. The related study indicates that the time delay of DT consists of computation and communication time costs and reveals the distinct impact of their sequence on DT latency. Moreover, the framework is further developed into a cross-platform prototype based on embedded systems, long-range wide area network (LoRaWAN), HTTP and MQTT protocols, restful web services, and IFC-based human-machine interface (HMI). The prototype is validated through different scenarios for bridge O&M, including drone-enabled bridge inspection, IoT-based bridge health-state monitoring, and decentralised dynamic evacuation. Secondly, an improved Prototypical Network (ProtoNet) is proposed for image-based bridge damage detection based on few-shot learning. It can work with only a few annotated examples, avoiding the tedious and labour-intensive data acquisition required by supervised learning. Feature embedding is “training free”, achieved through cross-domain transfer learning from ImageNet. The approach is explored on a public dataset through ablation studies and reaches over 94% mean accuracy for 2- way 5-shot classification via the pre-trained GoogleNet. Moreover, the proposed fine tuning methods are demonstrated with better performance than previous research. Finally, the approach is validated using real bridge inspection images, demonstrating its capability of fast implementation for practical damage inspection. Thirdly, a highly efficient framework is proposed for spatial damage assessment and DT synchronization based on point clouds in near real-time. The 2D damage is detected via DeepLabV3+ on pseudo grayscale images only from the point depth, avoiding the drawbacks of image and point cloud fusion. Then, 3D damage is separated through voxelization and converted into a lightweight binary matrix that can be further compressed losslessly for DT synchronization. The framework is validated via two case studies, demonstrating that the proposed voxel-based method can be easily applied to real-world damage with non-convex geometry instead of convex-hull fitting; FE and BIM models can be updated automatically through the framework. Fourthly, an automatic and unified deep learning (DL) framework is proposed for intelligent fault diagnosis (IFD) and health-state recognition based on time series by leveraging automated machine learning (AutoML) and data-level fusion. Uniaxial or triaxial signals can be reconstructed into 3-channel pseudo-images to satisfy the CNN input requirements and achieve data-level fusion simultaneously. Then, the model training, hyperparameter optimisation, and evaluation are carried out automatically. The selected model can be deployed on a cloud server or an edge device. Moreover, the framework can be extended by integrating multi-channel 1D-CNN architectures, validated using the data from a railway bridge's vibration-based monitoring (VBM) project. Multiple sensors' data-level and decision-level fusion performances are also compared and analysed. Fifthly, a knowledge graph (KG) schema based on bridge structure and maintenance reports is proposed. Then, graph data mining is explored on the established KG by leveraging large language models (LLMs) and graph neural networks (GNNs). The models trained from graph contextual similarity can identify node layer information and provide maintenance recommendation for unsolved defects from the existing options. Moreover, an intact workflow integrating the proposed KG schema and data mining approaches is designed for maintenance routine practices. Finally, a preliminary bridge DT system is developed based on the above outcomes. This research addresses some challenges encountered for bridge DT implementation in practical O&M and has paved the way for a more efficient, resilient, and intelligent bridge DT system. As such, it offers excellent potential to achieve generalised DT applications in smart and practical bridge maintenance.

Item Type: Thesis (PhD)
Date Type: Completion
Status: Unpublished
Schools: Engineering
Uncontrolled Keywords: 1)Digital Twin 2)Bridge Maintenance 3)AIoT-informed Communication 4)Few-shot Detection 5)Point-cloud Synchronization 6)Maintenance-oriented Knowledge Graph
Date of First Compliant Deposit: 1 February 2024
Last Modified: 01 Feb 2024 14:09

Actions (repository staff only)

Edit Item Edit Item


Downloads per month over past year

View more statistics