Data-driven pressure estimation and optimal sensor selection for noisy turbine flow with blocked clustering strategy

The design and control of turbomachinery require a wealth of spatiotemporal data. Thus, the low-cost and robust estimation of global aerodynamics from extremely limited data and noisy measurements is an important problem. This paper describes a data-driven approach to estimate the full-field pressure distribution of a turbine cascade flow in combination with sparse-distributed sensor measurements. For the offline library building and online reconstructing, the reduced-order model based on standard proper orthogonal decomposition (POD) and least squares approximation, and sparse representation based on overcomplete dictionary and L1 norm minimization are leveraged. To enhance the reconstruction accuracy and robustness with noisy measurements and varied sensor selections, a novel blocked K-means clustering strategy is developed to reconstruct the global flow field through the superposition of multiple local clusters. The statistical results indicate that sparse representation outperforms gappy POD in high-noise measurement environments due to its superior noise robustness and effective feature selection. By applying the proposed blocked clustering strategy, the accuracy and robustness of sparse estimation are significantly improved. The mean square error of gappy POD can be reduced by 9.86% for pressure reconstruction at 90% span of the turbine blade. Sparse representation produces excellent robustness enhancement when the noise intensity exceeds 0.3. Overall, the local reconstruction framework developed in this paper exhibits outstanding advantages in reconstruction accuracy and robustness.