Machine learning in flight parameter-based structural load prediction: A review and framework proposal
Lei Huang, Ying Zuo, Cong Guo, Bo Wang, Kuo Tian
Abstract
Flight Parameter-based Structural Load Prediction (FP-SLP) is a key technology for Structural Health Monitoring (SHM) and remaining life estimation of in-service aircraft and flight vehicles. With the continuous development of Machine Learning (ML), FP-SLP has gained powerful assistance. However, in engineering application of FP-SLP, ML technology faces bottlenecks such as fragmentation and data scarcity, which severely constrain its further development and application. This paper aims to conduct a systematic review of ML research and applications in FP-SLP related fields, establish a new advanced technical framework and explore future development directions, so as to provide references for theoretical research and engineering practice. In this paper, we first clarify the fundamental concepts of FP-SLP and describe the general workflow from data acquisition, surrogate modeling to online prediction, and comprehensively analyze existing challenges in engineering practice. Subsequently, a systematic review of ML applications of in FP-SLP is presented, concentrating on four core problems: efficient flight data preprocessing, high-precision surrogate modeling, model tuning and validation, and model deployment and maintenance. The advantages and effectiveness of advanced ML techniques in overcoming the existing challenges in FP-SLP are demonstrated. Based on various state-of-the-art ML techniques, this paper proposes a new advanced framework for FP-SLP, which provides a paradigm for subsequent research and engineering applications. Meanwhile, an original industrial-grade dataset, Aircraft structural Load Benchmark dataset (AirLoadBench), derived from real flight testing, is constructed and made public for the comparative research needs in the FP-SLP field. Finally, the AirLoadBench dataset is applied to evaluate various ML models in terms of prediction accuracy and training time. The results show that XGBoost model and XGBoost-based Ensemble learning (XGBoost-Ens) model demonstrates significant advantages among numerous models, outperforming others in both prediction accuracy and training efficiency.