PSI - Issue 80

Sadjad Naderi et al. / Procedia Structural Integrity 80 (2026) 77–92 Sadjad Naderi et al. / Structural Integrity Procedia 00 (2025) 000–000

79 3

system knowledge, and ultimately operational wisdom. In turn, this evolving digital intelligence should guide inspection priorities, and enable timely decision-making. DTs can be conceptualised as a four-dimensional engineering construct. Horizontally , they must integrate physical assets/sensor networks with data pipelines, inference engines, and computational models – a core focus of many recent frameworks (Chia et al. 2024). Vertically , each module demands progressive development, from advanced sensing and signal processing to hybrid AI–physics approaches and reduced-order modelling techniques (Tuegel et al. 2011). Temporally , DTs must maintain real-time or near-real-time synchronisation between physical and digital layers to ensure responsiveness under dynamic operating conditions (Sisson, Karve, and Mahadevan 2022). Uncertainty , the fourth dimension, must be quantified and propagated across all layers – encompassing sensor noise, model-form uncertainty, and material/environmental variability (Thelen et al. 2023). Despite substantial progress in each dimension, critical challenges remain. Scalable architectures remain rare, and many current implementations are only loosely coupled, lacking full bidirectional feedback and dynamic adaptability. Due to high computational cost, real time probabilistic simulations are constrained, and the integration of high-fidelity physics-based simulations to address data scarcity and enhance the generalisability of AI models remains limited. Furthermore, in-service validation, particularly in fatigue-critical structures, has been scarcely demonstrated. Collectively, these limitations hinder the operational deployment of DTs for structural integrity monitoring, crack evolution tracking, and reliable fatigue life prediction. This study develops and experimentally validates a probabilistic DT framework for fatigue crack prognosis in aerospace structures using Bayesian inference. Fig. 1 provides an overview of the framework and outlines how the current study fits within the broader DT architecture, including prior developments and future directions. The focus here is on the prognosis module, building on a previously developed passive sensing system based on higher harmonic analysis for real-time crack detection and quantification (Pan, Khodaei, and Aliabadi 2025). The sensing setup includes a multi-channel oscilloscope collecting waveform data from sensors placed on the test specimen. These data are used for real-time monitoring and quantification of crack growth. Two datasets are used: an offline/historical dataset from previous fatigue tests, and an online/observed dataset collected during the current experiment. The historical data are used to pre-train a model, which is then fine-tuned using the online data through a physics-informed Dynamic Bayesian Network (DBN) (Murphy 2002). This allows for probabilistic estimation of RUL, material properties, effective initial flaw size ( )(Morse, Khodaei, and Aliabadi 2019) – a latent parameter representing manufacturing-induced defects rather than the observed crack size detected by the diagnostic system. While the diagnostic stage provides early quantification of cracks, the prognosis module extends this capability by predicting the entire crack growth trajectory, from the to the critical failure point, under quantified uncertainty. To benchmark performance, the DBN-based approach is compared to a purely data-driven Gaussian Process Regression (GPR) model (Wang 2023). This comparison illustrates the trade-offs between physics-informed models that embed fatigue mechanics knowledge (DBN) and purely data-driven approaches that rely solely on statistical learning from observed data (GPR). By evaluating both, the study highlights the value of embedding domain knowledge for extrapolative tasks such as inference and long-term prognosis. While the present work focuses on the prognosis module, the framework is designed to be extensible. For example, although feedback-based mission and operation optimisation is not yet implemented, the architecture supports integration of such decision-making functionality in future developments. Overall, this study establishes a foundation for a closed-loop, adaptive DT system for structural health management.