Updating structural wind turbine blade models via invertible neural networks

Abstract

Wind turbine rotor blades are huge and complex composite structures that are exposed to exceptionally high loads, both extreme and fatigue loads. These can result in damages causing severe downtimes or repair costs. It is thus of utmost importance that the blades are carefully designed, including uncertainty analyses in order to produce safe, reliable, and cost-efficient wind turbines.

An accurate reliability assessment should already start during the design and manufacturing phases. Recent developments in digitalization give rise to the concept of a digital twin, which replicates a product and its properties into a digital environment. Model updating is a technique, which helps to adapt the digital twin according to the measured characteristics of the real structure. Current model updating techniques are most often based on heuristic optimization algorithms, which are computationally expensive, can only deal with a relatively small parameter space, or do not estimate the uncertainty of the computed results.

This thesis’ objective is to present a computationally efficient model updating method that recovers parameter deviation. This method is able to consider uncertainties and a high fidelity degree of the rotor blade model. A validated, fully parameterized model generator is used to perform a physics-informed training of a conditional invertible neural network. This network finally represents a surrogate of the inverse physical model, which then can be used to recover model parameters based on the structural responses of the blade. All presented generic model updating applications show excellent results, predicting the a posteriori distribution of the significant model parameters accurately.