On the application of a long-short-term memory deep learning architecture for aircraft dynamic loads monitoring
For aircraft, the contemporary approach to Structural Health Monitoring (SHM), known as Prognostics and Health Management (PHM), is driven by condition-based maintenance and structural health prognosis, where pre-emptive planning of maintenance and procurement of parts will allow for maintenance and sustainment practices to be optimized, offering major beneficial fiscal and safety implications for both the civilian and defense aerospace sectors. Despite some current reluctances in industry, AI-based technologies, such as, deep learning, may have a major role to play in the contemporary aircraft PHM system. With this in mind, this paper aims to serve as an important contribution to their broader adoption at this critical time in aircraft SHM, showing how the relatively simplistic implementation of the state-of-the-art in deep learning can change the entire potential capability of the aircraft SHM system. The Bidirectional Long Short-Term Memory (BiLSTM) deep neural architecture is introduced as an indirect Multi-Input Single-Output (MISO) modeling strategy for poorly conditioned (weak MISO coherence) load predictions. It is shown that a single fixed-architecture BiLSTM strategy can be used as a general purpose strategy for MISO modeling of the loads, with a prediction accuracy that is of an unprecedented levels of fidelity and robustness/consistency. More specifically, the previous use of more traditional strategies, such as, regression models and even standard artificial neural networks, has only been viable in terms of yielding a statistically reasonable distribution of extrema (from generally very poor MISO time-series models). Hence, in this research campaign (ongoing for the past four years), the introduction of the deep learning strategy has changed the paradigm by which the weakly coherent MISO airframe loads are modeled, where the high-fidelity point-to-point time-series predictions has introduced the viability of contemporary fatigue assessment algorithms.
