Temporal hierarchical modal neural network for structural strain prediction

Abstract Accurate in-flight structural strain prediction is critical for aircraft design, certification, and structural health monitoring. Traditional physics-based methods, such as finite element analysis or modal analysis, offer interpretable and robust frameworks to estimate structural strain or displacements from sparse measurements data. Nevertheless accurate yet representative in-flight dynamic loading requires extensive computational resources and their complex non-linear behaviors remains challenging. In contrast, learning-based approaches have shown promising results for structural strain prediction, but their performance remains limited by the learning domain spanned by the available data, and they lack interpretability. To take advantage of learning based approaches while retaining the interpretability and physical basis of traditional approaches, this work introduces three novel hybrid deep learning architectures that integrate modal information into neural network. After training on a frugal dataset of sparse strain measurement data, these architectures can accurately predict the strain field of an aeronautical structure from onboard instrumentation only. This enables structural strain virtual sensing on any new flight. In particular, the Temporal Hierarchical Modal Network (THM-net) relies on an original structuring of the learning process, which allows the contribution of structural modes to the structure’s response to be considered hierarchically. This architecture demonstrates high performances, maintains robustness to noise and mode shape uncertainty, while requiring limited measurements or structural information. These results highlight the potential of such approaches to bridge the gap between physics-based modeling and data-driven learning, and pave the way to the development of new hybrid architectures dedicated to in-flight structural dynamics monitoring of aircraft.