An efficient surrogate model for prediction of stress released distortion in large blade machining

Abstract

Deformation during the machining of large turbine blades based on a forged blank is always inevitable and unpredictable because of the individual difference in residual stress of the blanks. Compensation and control of the machining deformation is of great challenge since non-destructive measurement and accurate modeling of the distributed residual stress of complex surfaces are unavailable. To this end, this paper constructs a novel data-driven model based on the U-Koopman neural operator, which is an improvement of the Koopman neural operator to describe the relationship between the deformation after the current cutting stage and that of the next cutting stage. To avoid expensive experiments and tests, the finite element method is utilized to simulate the continuous multi-processes machining of large blades, which contains residual stress generation during forging process and deformation generation during cutting process, and then construct the dataset for model training. Cross-validation is implemented to verify the superior generalization ability of the proposed model over the benchmark models and the effectiveness of the related improvements of the model based U-Koopman neural operator. The results of case study show that the proposed model can predict and compensate for the deformation in-process and improve the machining accuracy and efficiency of large blades.