A pyramid-style neural network model with alterable input for reconstruction of physics field on turbine blade surface from various sparse measurements

Zuobiao Li, Fengbo Wen, Chenxin Wan, Zhiyuan Zhao, Yuxi Luo and Dongsheng Wen

Energy, 2024, vol. 308, issue C

Abstract: Data from turbine cascade experiments typically exhibits low spatial–temporal resolution, along with inevitable noise and local data missing. This paper aims to establish a super-resolution model to reconstruct the complete pressure and temperature fields on the blade surface from limited observations, using the deep learning method. Depending on the three-dimensional geometric complexity of the blade, the required cross-sectional data varies, resulting in input data of different sizes. Conventional surrogate models typically confine themselves to a single input format, leading to limited compatibility with diverse inputs. A pyramid-style model with four optional input ports is proposed, designed to accommodate data with different numbers of span-wise sections. Three training strategies have been evaluated, where distributed training has been proven to be more time-effective and flexible while maintaining high prediction precision, compared to holistic and conventional training. Generally, the average relative error over the whole dataset falls below 0.18%, the average peak signal-to-noise ratio exceeds 52 dB, and the average structural similarity index measurement surpasses 0.999. As the number of span-wise sections and the amount of information in the input increase, the overall performance shows a consistent upward trend. Robustness analyses are conducted by applying artificial noises of varying intensities. The results indicate that the model can tolerate noise intensities of up to 5% with a satisfactory reconstruction accuracy. In addition, the model is quite sensitive to the measurement data noises in complex flow regions such as expansion waves, suggesting that more intensive measurements should be targeted to those regions when conducting cascade experiments. The proposed method satisfies the intended objectives and provides an idea for future applications in digital twin platforms.

Keywords: Gas turbines; Supervised learning; Super resolution; Distributed training; Robustness analysis (search for similar items in EconPapers)
Date: 2024
References: View references in EconPapers View complete reference list from CitEc
Citations:

Downloads: (external link)
http://www.sciencedirect.com/science/article/pii/S0360544224026021
Full text for ScienceDirect subscribers only

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:eee:energy:v:308:y:2024:i:c:s0360544224026021

DOI: 10.1016/j.energy.2024.132828

Access Statistics for this article

Energy is currently edited by Henrik Lund and Mark J. Kaiser

More articles in Energy from Elsevier
Bibliographic data for series maintained by Catherine Liu ().