Unsupervised Learning for High-Fidelity Compression of Large Experimental Dataset: an Application to HPT Blade Tip Contouring

Authors

Francesco Aldo Tucci - Department of Mechanical and Aerospace Engineering, Sapienza University of Rome, Rome, Italy
Giovanni Delibra - Department of Mechanical and Aerospace Engineering, Sapienza University of Rome, Rome, Italy
Lorenzo Tieghi* - Department of Mechanical and Aerospace Engineering, Sapienza University of Rome, Rome, Italy
Alessandro Corsini - Department of Mechanical and Aerospace Engineering, Sapienza University of Rome, Rome, Italy
Sergio Lavagnoli - Turbomachinery and Propulsion Department, von Karman Institute for Fluid Dynamics, Sint-Genesius Rode, Belgium

Abstract

Most of the energy transition strategies in aircraft propulsion systems rely on increasingly complex design solutions meant to increase core engine efficiency. In this ambit, blade tip geometry of HPT rotors plays a crucial role in performance by controlling the tip leakage flows, reducing the rotor mixing losses and weakening thermal loads. Therefore, among all possible strategies, advanced rotor tip contouring is one of the most promising to refine current design strategies and improve the performance of high-speed turbines. However, the relationships between the geometric parametrization of the blade tip and the aerodynamic and thermal performance are highly non-linear and difficult to quantify with simple analytical formulations. In a view to infer and dissect the causal relationship among different tip geometries and unsteady aero-thermodynamic performance in HPT blading, this paper proposes a novel and generalized machine learning methodology to exploit large experimental data sets using unsupervised dimensionality reduction. The technique, advocating methods commonly used for image and sound recognition, is able to explore the functional relationships among performance and tip geometries. The novel machine learning procedure is validated on the non-sequential experimental dataset measured at von Karman Institute rotating turbine rig which operates in rainbow-rotor configuration: 48 blades divided in 7 sectors that share the same baseline rotor geometry but have different tip designs (optimized squealer-like rotor tips geometries). The novel approach, first, involves an automatic Python routine, based on Concave Hull Algorithm, to automatically identify the three-dimensional tip geometries. Dimensionality reduction is, thereafter, achieved by the concerted use of Principal Component Analysis (PCA) and Auto Encoder (AE) trained to provide latent representations of the input data. Those techniques are primarily used in image denoising, and to keep the accuracy during the training phase the algorithms are fed by different types of data (e.g., in image recognition it is common to feed VAE with cropped, rotated, and resized images). The latent representations obtained by PCA and AE define the input / output feature sets used with two artificial neural networks (ANNs) trained to link geometries to performance and vice versa. The latent layers predicted by ANNs are then decoded and the accuracy of the results with respect the original data is evaluated both on a pixel-by-pixel basis (with an overall error of less than 7%) and by measuring the Structural Similarity index (SSIM), which evaluates the similarity between two images and varies between -1 and 1, where 1 indicates perfect similarity. A SSIM of 0.97 is obtained.