Machine Learning Based Approach to Improve Low-Fidelity Predictions for a Compound Helicopter

Abstract

A novel machine learning approach is introduced, with the goal of minimizing the number of points required to correct a low-fidelity model using sparse high-fidelity data. The method is applied to a low-fidelity comprehensive trim analysis of a compound helicopter with three degrees of control redundancy: main rotor speed, auxiliary thrust, and stabilator setting. The final low-fidelity correction model applies small changes to the power requirement and main rotor trim control predictions to more closely match the high-fidelity data. To reduce the computational time, and labor cost of querying the high-fidelity data, the algorithm prioritizes data acquisition by iteratively selecting the data where the error is expected to exceed the model tolerance by the greatest margin. The model is trained until the error model anticipates a scaled accuracy of 5% at the minimum power region at each flight speed, with no more than 20% error across the entire trim envelope. For the low-fidelity correction model, this tolerance is met through training with data from 89 trim states, which is a 79% reduction from the 428 trim states required to obtain comparable accuracy for a purely data-driven machine learning approach. Evaluation on a testing data set yields a rate of 95-98% of testing points actually falling within the error tolerance goals, compared to 80-93% for the purely data-driven model.