Uncertainty-aware deep kernel learning: An end-to-end approach for crack localization in turbine blades

Area of Science:

• Aerospace Engineering
• Mechanical Engineering
• Machine Learning

Background:

• Early crack detection in jet turbine blades is crucial for aerospace structural health monitoring.
• Modal frequency data offers a promising avenue for non-destructive crack localization.

Purpose of the Study:

• To develop and evaluate an uncertainty-aware Deep Kernel Learning (DKL) framework for crack localization in jet turbine blades.
• To assess the DKL model's performance and uncertainty estimation capabilities under various noise and data shift conditions.

Main Methods:

• An end-to-end DKL framework integrating a deep residual feature extractor with a Sparse Variational Gaussian Process (SVGP) model was employed.
• A finite element simulation dataset with 901 samples was generated, varying crack positions and extracting modal frequencies.
• Robustness was tested against Gaussian noise, synthetic dataset shifts, and realistic non-uniform stochastic perturbations.

Main Results:

• All models showed high accuracy on clean data; however, under stochastic perturbations, Deep Ensemble yielded the lowest prediction error.
• The proposed DKL model demonstrated competitive accuracy while providing more conservative and stable uncertainty estimates compared to other methods.
• The DKL framework offers a balanced trade-off between predictive accuracy and uncertainty awareness.

Conclusions:

• The uncertainty-aware DKL framework shows potential for reliable crack localization in turbine blades.
• The model's ability to provide stable uncertainty estimates is valuable for structural health monitoring applications.
• Further validation with real-world sensor data is necessary to confirm practical applicability.