Research on damage prediction methods for service turbine blades based on meta-learning

Abstract Accurate prediction of microstructural damage evolution in in-service aero-engine turbine blades is difficult because labeled service data are scarce and blades from different production batches exhibit different initial microstructures. This study applies a model-agnostic meta-learning framework with a long short-term memory base learner (MAML-LSTM) to small-sample damage prediction of K403 nickel-based superalloy turbine blades. A batch-independent damage variable was defined using the γ′-precipitate morphology of the tenon as a blade-specific reference, and 12 spatially distinct blade regions were formulated as related meta-learning tasks according to differences in local temperature, stress and degradation kinetics. To prevent augmentation-induced information leakage, all image patches and noise-augmented derivatives generated from the same parent SEM image were kept in the same task and on the same side of each support/query split. Across 10 independent runs, the proposed model achieved a test-set MAE of 0.0162 ± 0.0020, MSE of (3.80 ± 0.90) × 10⁻⁴, RMSE of 0.0195 ± 0.0023, and R² of 0.9786 ± 0.0064. On the validation tasks, MAE and MSE were reduced by 57.82% and 83.82%, respectively, relative to a tuned LSTM baseline; the improvement remained evident against classical small-sample regressors including SVR and Gaussian-process regression. Ablation and support-size experiments show that both meta-learning and controlled data augmentation contribute to accuracy, with the largest advantage occurring at the smallest support sizes. Validation on three additional datasets indicates transferability across related regression tasks, although the small sizes of the external datasets limit claims of broad cross-domain generalization. The framework therefore provides a data-efficient route for microstructure-based blade damage assessment and can support, rather than replace, physics-based life assessment and maintenance decision procedures.