This thesis presents a machine learning (ML) based method for sensing both the shape and applied distributed loads of a highly flexible wing (the Pazy wing) using measured strain data. The predicted shape and load distribution are then integrated with an aerodynamic model to estimate the rigid angle of attack (RAoA).
The research employs a practical approach, utilizing relatively simple static ground tests with a few concentrated loads to train the ML model, which subsequently predicts aerodynamic distributed loads and wing deformation during wind tunnel tests.
A comprehensive workflow for designing training experiments and evaluating results is detailed, complemented by a well-structured data processing and training pipeline. The concept is validated using ML models initially trained on a finite element (FE) structural model of the wing, simulating loads that produce geometrically nonlinear deflections up to 50% (275mm) of the wing’s 550mm span.
The performance of the load and shape prediction models is then evaluated using actual wind tunnel test data with deflections up to 100mm, achieving maximum errors below 9.1%. The generalization and extrapolation capabilities of the models are assessed and demonstrate robust performance beyond the training domain.
Finally, the ML models are employed to determine the wing’s aerodynamic characteristics and estimate the rigid AoA in wind tunnel tests ranging from -10° to 10°, with a maximum error of 5.2% (0.6°) within the model’s training range.
This study successfully establishes a comprehensive methodology for predicting in-flight shape and loads using ground-test-trained ML models, providing an effective approach for estimating RAoA for highly flexible structures.
Light refreshments will be served before the lecture
This thesis presents a machine learning (ML) based method for sensing both the shape and applied distributed loads of a highly flexible wing (the Pazy wing) using measured strain data. The predicted shape and load distribution are then integrated with an aerodynamic model to estimate the rigid angle of attack (RAoA).
The research employs a practical approach, utilizing relatively simple static ground tests with a few concentrated loads to train the ML model, which subsequently predicts aerodynamic distributed loads and wing deformation during wind tunnel tests.
A comprehensive workflow for designing training experiments and evaluating results is detailed, complemented by a well-structured data processing and training pipeline. The concept is validated using ML models initially trained on a finite element (FE) structural model of the wing, simulating loads that produce geometrically nonlinear deflections up to 50% (275mm) of the wing’s 550mm span.
The performance of the load and shape prediction models is then evaluated using actual wind tunnel test data with deflections up to 100mm, achieving maximum errors below 9.1%. The generalization and extrapolation capabilities of the models are assessed and demonstrate robust performance beyond the training domain.
Finally, the ML models are employed to determine the wing’s aerodynamic characteristics and estimate the rigid AoA in wind tunnel tests ranging from -10° to 10°, with a maximum error of 5.2% (0.6°) within the model’s training range.
This study successfully establishes a comprehensive methodology for predicting in-flight shape and loads using ground-test-trained ML models, providing an effective approach for estimating RAoA for highly flexible structures.
Light refreshments will be served before the lecture
