Inverse Method of Centrifugal Pump Blade Based on Gaussian Process Regression

Abstract

The inverse problem is always one of the important issues in the field of fluid machinery for the complex relationship among the blade shape, the hydraulic performance, and the inner flow structure. Based on Bayesian theory of posterior probability obtained from known prior probability, the inverse methods for the centrifugal pump blade based on the single-output Gaussian process regression (SOGPR) and the multioutput Gaussian process regression (MOGPR) were proposed, respectively. The training sample set consists of the blade shape parameters and the distribution of flow parameters. The hyperparameters in the inverse problem models were trained by using the maximum likelihood estimation and the gradient descent algorithm. The blade shape corresponding to the objective blade load can be achieved by the trained inverse problem models. The MH48-12.5 low specific speed centrifugal pump was selected to verify the proposed inverse methods. The reliability and accuracy of both inverse problem models were confirmed and compared by implementing leave-one-out (LOO) cross-validation and extrapolation characteristic analysis. The results show that the blade shapes within the sample space can be reconstructed exactly by both models. The root mean square errors of the MOGPR inverse problem model for the pump blade are generally lower than those of the SOGPR inverse problem model in the LOO cross-validation. The extrapolation characteristic of the MOGPR inverse problem model is better than that of the SOGPR inverse problem model for the correlation between the blade shape parameters can be fully considered by the correlation matrix of the MOGPR model. The proposed inverse methods can efficiently solve the inverse problem of centrifugal pump blade with sufficient accuracy.