Get started
Work

Ultrasone Geleide Golf Structurele Gezondheidsmonitoring: Een Probabilistisch Deep Learning Perspectief

Year: 2025
Type: dissertation
Abstract: Intro: Structural Health Monitoring (SHM) is a vital technology in engineering design and operational management focused on ensuring structural efficiency and safety. It plays an essential role in extending structural service life, minimizing maintenance costs, and preventing catastrophic failures. A central objective of SHM is the early detection of structural anomalies before they lead to critical damage, thereby enabling preventive inspection strategies. Ultrasonic Guided Waves (UWGs), which are boundary-guided, exhibit low attenuation over long distances and are highly sensitive to damage and defects—even minor damage can lead to observable changes in wavefield features. As a result, UGW monitoring has become one of the most widely adopted non-destructive techniques in SHM. Deep learning has recently emerged as a popular research method across various SHM applications and is expected to play an increasingly pivotal role in the future. By employing deep learning models for feature extraction, anomaly detection, and predictive analysis of UGW data from damaged structures, SHM systems can achieve higher levels of automation, improved real-time monitoring, and enhanced capacity for managing complex data. Why: Despite recent advancements, SHM technologies that integrate traditional deep learning with UGWs continue to face several significant challenges: (i) effectively extracting damage-sensitive features from UGWs; (ii) traditional deep learning models are unable to provide statistical grounded confidence estimates for their predictions; (iii) the difficulty of acquiring high-quality labeled data in many practical settings; (iv) limited generalizability of models trained in one domain or scenario when applied to others. To enable the practical deployment of UGW-based SHM systems, it is essential to address four interconnected challenges: feature reliability, uncertainty, label scarcity, and domain shift. How: To address the core challenges, this work proposes a systematic technical framework: feature engineering→probabilistic deep learning→probabilistic unsupervised transfer learning. Specifically, • Feature engineering: (a) Data level: A various UGW database is developed via numerical simulations and experimental measurements, encompassing four scenarios: delamination in composite beams, microcracks of 3D-printed auxetic honeycomb structures, surface defects in aluminum-stiffened panels, and corrosion in underwater steel pipes. (b) Feature level: A novel five-dimensional systematic evaluation criterion is introduced to quantitatively assess 15 damage-sensitive features across time, frequency, and time-frequency domains. • Probabilistic deep learning: Enhancements to the VGG-13 backbone with probabilistic layers to form three probabilistic deep learning models—Flipout probabilistic convolutional neural network (FPCNN), deep ensemble PCNN, and Bayesian PCNN—each integrating uncertainty-aware components and leveraging techniques like Flipout variational inference, Bayesian layers, and deep ensembles. These models provide not only diagnostic outputs but also 95% confidence intervals capturing both aleatoric and epistemic uncertainties. • Probabilistic unsupervised transfer learning: Two unsupervised transfer learning frameworks are proposed. The first, Multi-scale Adaptive Attention Transformer Domain Adaptation Network (MAAT-DAN), combines a multi-scale adaptive attention Transformer with adversarial feature alignment and maximum mean discrepancy regularization to achieve domain adaptation without target domain labels. The second, Deep-Ensemble Unsupervised Transfer-Learning (DE-UTL), integrates deep ensembles with Wasserstein distance to further improve the same label-free adaptation while also quantifying both types of uncertainty. Results: The constructed multi-dimensional feature evaluation framework provides a reference for processes of feature selection and evaluation. Furthermore, the proposed probabilistic deep learning framework demonstrates strong generalization capabilities by maintaining low prediction errors across diverse conditions, including different materials (composites, auxetic structures, aluminum alloys, steel pipes), environments (from air to water), and intersensor/intertemporal (from 2012 to 2013). (1) In feature evaluation, this work proposes five criteria—diagnostic capability, environmental stability, computational complexity, sensor configuration, and structural applicability—and applies them to classify 15 common features into four performance levels. (2) In the composite delamination dataset, the Flipout-based network achieves an optimal balance between training efficiency, uncertainty quantification effectiveness, and accuracy in localizing damage. (3) In 3D-printed auxetic structures, FPCNN consistently provides compression deformation predictions within the 95% confidence interval of total uncertainty. It is also capable of early identification of critical damage at the point of a sharp drop in envelope energy. (4) In metal structures, MAAT-DAN significantly outperforms conventional models in unsupervised cross-material regression and cross-temporal sensor classification tasks, showcasing its robustness in feature extraction, domain alignment, and stable generalization across domains. (5) In underwater structures, DE-UTL demonstrates high accuracy in unsupervised corrosion thickness diagnosis and uncertainty quantification under geometry-environmental transfer from a rectangular cross-section steel pipe in air to a square cross-section pipe in water. Overall, the experimental results confirm that the proposed guided wave-based probabilistic deep learning SHM framework achieves high accuracy and reliability in scenarios involving different materials, harsh environments, and task switching. It significantly reduces the need for labeled data and potentially improves engineering deployability. This work establishes a comprehensive technical roadmap for UGW SHM—from feature engineering to probabilistic inference, and unsupervised transfer learning—laying a solid theoretical and practical foundation for future intelligent, reliable, cross-scenario, and cost-effective SHM systems.
Author: Lu, Houyu
Language: English
Open Access status: green
OpenAlex ID: W7112447357
Cites:
Cited by:
Related to: