Damage detection in composite structures utilising machine learning and cross-domain adaptation
Rezazadeh, Nima (2026) Damage detection in composite structures utilising machine learning and cross-domain adaptation. Doctoral thesis, Birmingham City University.
Preview |
Text
Thesis_Nima Rezazadeh (Compressed).pdf - Submitted Version Download (3MB) |
Preview |
Text
A transfer learning approach for mitigating temperature effects on wind turbine blades damage diagnosis (Compressed).pdf - Submitted Version Download (1MB) |
Preview |
Text
Degradation mode identification and remaining useful life prediction via an interpretable CNN-BiLSTM framework (Compressed).pdf - Submitted Version Download (986kB) |
Preview |
Text
Domain-Adaptive Graph Attention Semi-Supervised Network for Temperature-Resilient SHM of Composite Plates (Compressed).pdf - Submitted Version Download (870kB) |
Preview |
Text
Systematic Review (Compressed).pdf - Submitted Version Download (652kB) |
Preview |
Text
WaveCORAL-DCCA A Scalable Solution for Rotor Fault Diagnosis Across Operational Variabilities (Compressed).pdf - Submitted Version Download (706kB) |
Preview |
Text
A novel interpretable domain adaptive framework for robust damage detection in composite structures under environmental variability (Compressed).pdf - Submitted Version Download (2MB) |
Abstract
Structural health monitoring of composite structures in realistic industrial settings is constrained by 3 persistent difficulties. First, labelled data describing damaged conditions are inherently scarce because deliberate damage is costly, time-consuming and often impractical. Second, sensor responses are strongly influenced by environmental and operational variability, such as changes in temperature and loading, which alter the measured signals even when the structural state is unchanged. Third, models that are trained under a particular set of conditions often fail to generalise when deployed under different conditions, so that performance degrades precisely when reliability is most needed. This thesis addresses these challenges by developing and evaluating a coherent set of machine learning frameworks that jointly tackle data scarcity, shifting operating conditions and the need for interpretable, computationally practical diagnosis of damage in composite laminates and a small-scale wind turbine blade.
The work begins by strengthening the data foundation through a suite of augmentation and dataset completion procedures tailored to structural health monitoring signals. Windowing is used to segment long records into multiple shorter samples in a manner that preserves temporal coherence. Spline-based interpolation across temperature is introduced to generate physically plausible signal variations between a limited number of measured operating points. A convolutional conditional variational autoencoder is designed to learn a latent representation of the signals that is explicitly conditioned on both damage state and temperature, and this representation is then used to synthesise new signals under unmeasured conditions whilst preserving damage-sensitive patterns in the time, frequency and time–frequency domains. In the feature space, Synthetic Minority Over-sampling Technique is employed to rebalance classes when damaged conditions are underrepresented. These components are used selectively to expand the effective training set without distorting the underlying physics of wave propagation and vibration.
The novel contribution of this thesis lies in the development of three complementary damage detection frameworks, each tailored to different regimes of domain shift and supervision availability, combined with interpretability and computational tractability as primary design constraints. The first framework is designed for mild shifts with a small number of labelled examples in the target condition. It introduces a novel integration of manifold learning based on uniform manifold approximation and projection with a capsule-based neural feature extractor and an instance reweighting strategy that downweights misleading source examples whilst upweighting informative ones, followed by an ensemble classifier. The second framework addresses multi-sensor, semi-supervised adaptation. It proposes a novel architecture that employs a graph attention-based encoder to model the spatial relationships between sensors and actuator–sensor paths, and aligns source and target feature distributions by combining adversarial training with statistical alignment of their means and covariances. The third framework is intended for severe domain divergence without any target labels. It presents a novel prototype-guided approach that uses wavelet time scattering to obtain deformation-stable and translation-invariant features, and then performs prototype-guided adversarial domain adaptation in which class prototypes are iteratively refined, target samples are pseudo-labelled under confidence and similarity constraints, and alignment is encouraged around these prototypes rather than around raw features, which helps to maintain class separation.
Interpretability and computational efficiency are treated as primary design requirements rather than afterthoughts. The proposed pipelines provide diagnostic views that include trajectories of samples in the learned manifolds, capsule-based feature activations associated with specific damage states, attention weights that reveal the relative contribution of each sensor, and the evolution of class prototypes during adaptation. All methods are implemented on central processing unit workstations with constrained hyperparameter searches in order to reflect realistic industrial computation budgets. Across two heterogeneous case studies, namely a small-scale glass fibre reinforced epoxy wind turbine blade tested under controlled temperature variation and a carbon-epoxy composite plate monitored by guided waves for multiple damage severities, the proposed frameworks consistently outperform strong base line methods in their respective regimes of shift and supervision. In conclusion, this thesis demonstrates that tailored domain adaptation strategies, when combined with physically informed augmentation and explicit interpretability mechanisms, can substantially improve the reliability and deployability of machine learning-based damage detection systems for composite structures under realistic operational variability. Taken together, the augmentation procedures and three domain adaptation frameworks form a practically deployable toolbox that mitigates data scarcity, accommodates changing environmental and operational conditions and delivers auditable damage detection for composite structures.
Actions (login required)
 |
View Item |
Tools
Tools
