Real-time identification of structural random dynamic loads using strain-based LSTM for Digital Twin applications

In such cases that accelerometers are impractical, the challenge lies in reconstructing interface acceleration loads in real time from sparse strain measurements, a prerequisite for sustaining high-fidelity Digital Twins subject to random vibration. To address such a task, there is in this paper a strain-based Long Short-Term Memory (LSTM) model for predicting acceleration load estimation with experimental verification of the model satisfaction with Digital Twin scenarios’ real-time and computational constraints. Upon learning the complex dynamic properties with the structural response time series, such a model is capable of giving correct and timely estimation of acceleration loads. Its effectiveness is illustrated through experimentation with the mock-up cabin of a rocket subjected to random and fixed-frequency vibration excitations, under such circumstances the proposed model outperforms among all the neural network configurations in terms of robustness and accuracy, which indicates LSTM-based methods is more suitable for real-time load identification and subsequently embedding into digital twin pipeline for structural health monitoring and predictive maintenance.