Fast High-Dimensional Parameter Optimization for Turbine Blade Manufacturing Using the Powerball L-BFGS Method Under Incomplete Measurements

Proper optimization of machining parameters can effectively improve production efficiency and precision in high-precision machining, such as turbine blade manufacturing. However, in industrial practice, current studies on the optimization of process parameters suffer from two drawbacks: the first is missing measurement values due to network latency and packet loss, while the second is the slow convergence rate of high-dimensional parameters to be optimized. This article proposes a prediction and optimization framework based on an accelerated limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) algorithm; the framework addresses the high-dimensional parameter optimization issues in high-precision machining processes, under incomplete sensor measurement scenarios. Incomplete measurements are imputed based on learned data distributions, significantly improving data utilization of the subsequent deep learning model trained after imputation. Then, to achieve fast high-dimensional parameter optimization, we introduce an accelerated L-BFGS algorithm, termed Powerball L-BFGS, with a Powerball function to optimize the search direction. To evaluate the effectiveness of the proposed framework, we conduct experiments on two series of aviation turbine blades (named Blade A and Blade B) with complex curved surfaces. Incomplete coordinate measuring machine data at key stages (i.e., blade root milling and comprehensive accurate milling) are used as inputs to predict and optimize the final geometrical errors. The proposed Powerball L-BFGS reduces the final optimization root mean square error (RMSE) to 0.053 mm for Blade A and 0.028 mm for Blade B; it decreases optimization iterations from 60 to 21 for Blade A and from 85 to 32 for Blade B, and achieving maximum process capability values of 5.23 and 2.39, respectively.