Abstract
An ultrasonic guided wave-based structural health monitoring system has potential applications in mainy domains such as, the oil and gas industry, civil engineering, and aerospace. However, there are some inherent challenges, such as the sensitivity of the Guided Waves (GW) to environmental and operational conditions (EOCs), defect(s) size and location, and sensor(s) placement. Therefore, the reliability of detection systems based on GW requires validation. Simulation tools are often used to study the impact of the above-mentioned factors. However, the computational burden associated to extensive simulation campaigns is excessive. To increase the computational efficiency, this work proposes a machine learning-based Digital Twin (DT) framework. More specifically, the DT framework employed in this paper comprises a linear dimensionality reduction algorithm and fully connected neural networks that work as a metamodel. The performance of the DT is evaluated on a simulation configuration of Aluminum with uncertainties in instrumentation and damage size. The simulation data required for training are obtained from CIVA simulation platform. The predicted signals from the DT are quantified using misfit-based criteria targeting amplitude and phase aspects based on time-frequency transformation. The assessment of the results suggests that DT has captured all the dynamics of the signals, and the predicted signals are in good agreement with the simulated ones. Furthermore, the developed DT has been employed to efficiently carry out the probability of detection study for reliability assessment and sensitivity analysis based on the propagation of uncertainties.
