Precision blade manufacturing: small-sample prediction and optimization using improved meta-learning and Particle Swarm Optimization

english

2025

volume 20, issue 3

machine learning   meta-learning optimization   optimization   Particle Swarm Optimization   precise manufacturing   PSO   small sample learning

Univerza v Mariboru, Fakulteta za strojništvo
(Obvezni izvod spletne publikacije)

Accurately predicting blade manufacturing deviations from limited experimental data remains challenging due to the complex nonlinear relationship between process parameters and resulting profile deviations in precision casting. To overcome the limitations inherent in traditional approaches and conventional machine learning methods, this study proposes a novel prediction and optimization framework specifically designed for small-sample scenarios, integrating enhanced meta-learning optimization with advanced Particle Swarm Optimization (PSO). We innovatively improve the model-agnostic meta-learning (MAML) algorithm by incorporating a dynamic loss function weighting strategy and a stochastic gradient descent with warm restarts (SGDR) learning rate mechanism, significantly mitigating overfitting and enhancing generalization performance. Additionally, we propose a process parameter optimization model utilizing an improved PSO algorithm with dynamic inertia and adaptive learning factors, designed to effectively navigate high-dimensional optimization landscapes. Experimental validation using orthogonal design data highlights pulling speed as the dominant factor influencing blade deviations (Pearson correlation coefficient (r = 0.67). The optimized parameters—low pulling speed (1.5 mm/min) and high pouring temperature (1530 °C)—achieve an 11.54 % reduction in blade deformation. The improved MAML-based prediction model demonstrates superior accuracy, achieving a mean absolute error (MAE) of 2.566 × 10-4 mm, representing a 21.7 % improvement over traditional Adam optimization methods, and exhibits robust predictive capability (R2 = 0.92) in small-sample contexts. This research not only delivers practical insights and precise parameter recommendations for complex blade manufacturing processes but also establishes a robust methodological framework applicable broadly to precision manufacturing domains characterized by limited data availability.

06.07.2026