| dc.description.abstract | Monitoring the systems or structures continuously for integrity and degradation is defined as structural health monitoring (SHM). Recent advances in SHM heavily rely on the application of diverse machine learning (ML) algorithms. SHM via ML algorithms has a range of applications across different fields such as civil, mechanical and aerospace infrastructure sectors. However, employing ML algorithms is highly
challenging due to data scarcity. In particular, supervised ML algorithms require labelled data for each potential damage scenario of the system or structure. Acquiring labelled data for these potential damage cases could uplift the objective of generating self-sufficient and well-trained algorithms. It is a challenging task to anticipate
potential damage scenarios of any structure and to obtain corresponding labels. Especially, while dealing with expensive structures such as skyscrapers or aircraft. Keeping the safety and economic factors in consideration there is a need for robust ML algorithms which can potentially use damage surrogates for training and can successfully detect the existence of various damage scenarios of the structure.
The primary objective of this thesis is to utilise numerical models and ML algorithms for generating surrogate data without the use of experimental damage data, particularly in the supervised setting. Initial data generation is accomplished with the aid of high-fidelity Finite Element (FE) models, which depict detailed and accurate representations of the physical systems or structures. However, incorporating FE models as surrogates may not effectively work for all types of damage scenarios and structures in training ML algorithms. In line with the research objective, the employment of algorithms such as clustering, and domain adaptation algorithms like transfer component analysis (TCA) and Joint Domain Adaptation (JDA) proved instrumental in addressing the challenge of data insufficiency.
Initially, a cluster-based approach of generating damage surrogates was implemented for effective damage identification on a Brake-Reuß beam structure. However, this method uses corresponding experimental damage states for projecting clusters.
To achieve the objective of not utilising experimental damage data, a domain adaptation methodology was employed to generate surrogate damage data utilising both numerical and experiential normal condition data. These surrogate damage data
(Numerical) facilitated training multiple classifiers for multi-class damage detection. This domain adaptation methodology is further extended for a localisation problem with damage cases such as saw-cuts on two homogenous Brake-Reuß beams.
With this strategy, the thesis mitigates the problem of the lack of training data for various damage classes. This study examines this methodology across structures of varying complexity through a range of structures including an experimental
beam, a laboratory wing box structure, and a real aircraft tailwing. Throughout this thesis, experimental structures undergo damage simulation involving loosened bolts and sawcuts from Brake-Reuß beam, wingbox laboratory structures and Piper 28 aircraft tailwings. Domain adaptation methodology obtained promising results (on experimental data) in training multiple classifiers including neural networks,
KNN classifiers, and ensemble classifiers. Damage detection is achieved on all three structures, with the localisation achieved specifically on the beam and tail wing structures. Moreover, this research underscores the importance of considering
negative transfer in transfer learning. Additionally, it highlights the significance of feature and hyperparameter selection (for domain adaptation) to mitigate negative transfer effects.
Finally, conclusions of the above research are provided. In addition, this thesis concludes by outlining potential areas for future work. Key suggestions include exploring the integration of Generative Adversarial Networks (GANs) to generate
synthetic data with variability, enhancing domain adaptation techniques to consider environmental effects on structures, and developing automatic feature selection methods using Genetic Algorithms (GAs) and Convolutional Neural Networks (CNNs).
Additionally, incorporating digital twins alongside domain adaptation is proposed to improve data quality and predictive accuracy. These advancements could lead to improved generalisation of SHM methodologies across different operating conditions, environments, or structural configurations, paving the way for more robust structural health monitoring solutions. | en_GB |
