Machine Learning for Structural Health Assessment under Uncertainty: with applications in Wind Energy

Abstract

The engineering sub-discipline of Structural Health Monitoring (SHM) promises that actionable insights on the current and future structural condition of monitored structures (e.g., wind turbines, bridges or critical components such as bearings) can be achieved through sensing and the subsequent inference of suited computational models. The main tools of SHM are system identification and computation, with the possible incorporation of physics-based models. Wind energy generation is an important factor in directly mitigating climate change by facilitating the reduction of energy production-related carbon emissions. The viability of wind energy projects, hinges upon the ratio of the expected long-term power produced, to the cost of installation, operation, repair, and maintenance of wind turbines. By forecasting and detecting damages and potential degradation of wind turbines in a timely and robust manner, a core target of SHM, repair and maintenance-associated downtimes can be reduced, thus improving the long-term viability of wind energy infrastructure. Currently, available SHM techniques underperform when dealing with very high-dimensional and high-volume data and often under-deliver due to simplifications for computational convenience of probabilistic and model-form assumptions underlying them. Through the course of the last two decades, Machine Learning (ML), and in particular Deep Learning, routinely proves to be a superior approach to flexible function approximation when adequate amounts of data are available. Moreover, in the last decade, scalable and effective techniques have emerged for the representation of uncertainty directly from data via the use of deep neural networks. This dissertation introduces these deep learning tools, to SHM problems, with a particular focus on wind energy applications. In data-driven methods, and in particular, in deep learning, the underlying structure of problems is often neglected, since the main expectation is that these models discover the hidden structure directly from data. This approach often introduces spurious correlations between variables and negatively affects generalization. In addressing this issue, this thesis proposes a technique that allows for the flexible exploitation of known, or partially known, relations of entities involved in a problem (relational inductive biases), termed Graph Neural Networks (GNNs). The first application introduced in this thesis in chapter 5, is the application of variational Bayesian deep neural networks for the distribution of cross-section fatigue estimates available in finite element simulation meshes. The proposed technique is useful in estimating the distribution of fatigue estimates on turbine blade cross-sections, from coarse SCADA information and fatigue simulations, to enable long-term fatigue accumulation prediction under uncertainty. A problem in wind-energy related SHM, where it is of use to incorporate the problem structure, is the modeling of the statistics of operational data of turbines in a farm, where encoding the relative position of turbines in the learning algorithm is an mportant relational inductive bias. This case is treated in chapter 6. The technique is demonstrated in real and simulated farm data, demonstrating generalization in novel farm geometries for the simulated data. It is shown that the proposed technique comprises a combination of other recently proposed approaches to learning on data from stochastic processes, and shows good generalization on a synthetic benchmark example for one-dimensional Gaussian process regression. In chapter 7, a supervised learning problem where the incorporation of geometric relations plays a key role is presented. The problem treats the fusion of data from an arbitrarily positioned set of sensors and, in particular, sensors deployed on a beam of circular cross-section. The relative position of the sensors needs to be incorporated in the learning procedure since in order to infer the position of a defect, it needs to be taken into account jointly with the sensor readings. The developed technique performs well with a relatively small number of training examples and incorporates variational Bayesian neural network layers that allow for representing the uncertainty in the learned model. In chapter 8, predictive data-driven models are developed for addressing the ultimate step of the SHM hierarchy, i.e., the problem of remaining useful life (RUL) estimation. The models are trained on time series corresponding to synthetic and real data on degradation of critical structural/industrial components, namely bearings. Through the use of graph networks, predictive models that operate on irregularly sampled data, which do not process the time series in a sequential manner are developed. Although the actual bearings where this method was employed do not correspond to bearings from wind turbines, the developed techniques are straightforwardly transferable to real use-cases. The proposed GNN techniques cover supervised (predictive tasks, such as localization), unsupervised (deep latent variable modeling for the distribution of operational data), and semi-supervised learning (i.e., data imputation for operational data). The dissertation concludes in chapter 9, where further applications and possible extensions of this work are discussed.