An Efficient Quantification Method Based on Feature Selection for High-Dimensional Uncertainties of Multistage Compressors

Multistage compressors are widely used in aero-engines, gas turbines, and industrial compressors. However, many inevitable geometric/condition uncertainties will lead to low manufacturing qualified rate and severe performance degradation of multistage compressors. There might be hundreds of uncertainty variables simultaneously, so the "curse of dimensionality" problem has seriously constrained the uncertainty quantification (UQ) study in multistage compressors. In this paper, a feature selection method based on interpretable machine learning and bidirectional search is developed to reduce the dimensionality in a multistage compressor UQ study. The method is applied to a three-stage compressor and reduces the uncertainties dimensionality from 292 to 41 and 66 for mass flow rate and efficiency prediction. Consequently, the required sample size is reduced from 1674 to 160 with the model accuracy almost unchanged. For the first time, this study realizes the UQ modeling study on the high dimensional problem of a multistage compressor. In addition, significant stage-by-stage uncertainty propagation is found in the compressor. The last stage has the most significant efficiency deviation, which is significantly affected by the geometric uncertainty of the first two stages. Therefore, traditional studies, which usually simplify the multistage compressor UQ problem to a single cascade/row level, may seriously underestimate the uncertainty effect due to the neglect of stage-by-stage propagation. This study not only reveals the necessity for direct UQ study of multistage compressors but also reduces the computational cost, which lays a foundation for the uncertainty control of multistage compressors in engineering practice.