Robust acceleration schedule design for gas turbine engine using multilayer perceptron network with adaptive sample class weighting

Abstract

The acceleration schedule is crucial for generating acceleration control references for the gas turbine engine (GTE) and ensuring optimal acceleration performance. However, under harsh operating conditions, GTEs may encounter difficult-to-diagnose pressure sensor faults, which lead to inaccurate control references and decreased acceleration performance. This paper proposes a data-driven robust acceleration schedule (RAS) design method to tackle this issue, including fault data augmentation and adaptive sample class weighting (ASCW). Fault data augmentation generates multiple pressure fault sample classes from a normal acceleration schedule dataset. Due to the redundant pressure sensor configuration, the RAS can reconstruct these classes and generate accurate control references when a single pressure sensor fails. The ASCW employs a multilayer perceptron network to reconstruct the normal and fault sample classes accurately. Proportional integral regulation adjusts their weights during training to ensure balanced reconstruction precision. Simulation cases were conducted to verify the effectiveness of the RAS under actual GTE conditions, including engine-model mismatches, performance deterioration, flight envelope, and measurement uncertainty. The results demonstrate that the RAS ensures superior acceleration performance of GTEs across the full flight envelope in both normal and single pressure sensor fault scenarios. Additionally, the ASCW achieves reconstruction precision of 0.068 %, 0.080 %, 0.080 %, 0.082 %, and 0.077 % for normal and fault sample classes, respectively, prioritizing the precision of the normal sample class and balancing the precision of fault sample classes.