Integrating Digital Twin and Random Survival Forests for Predicting Fatigue Behavior in Materials

Fatigue failure poses a significant threat to the safety and reliability of mechanical systems, often leading to catastrophic and unpredictable outcomes. Traditional deterministic models and reactive maintenance strategies are inadequate for modern industrial applications, where assets are expected to operate continuously under varying loads and environmental conditions. The challenge is compounded by the prevalence of censored data, where the exact time of failure is unknown for many components. To address this critical need, this paper proposes a novel framework for a digital twin dedicated to prognostics and health management (PHM). Our methodology leverages the power of survival analysis, a statistical branch uniquely suited to modeling time-to-event data with censoring. We utilize a Random Survival Forest model to establish a probabilistic relationship between operational parameters and a component's remaining useful life (RUL). The model is trained on a comprehensive dataset, integrating physically-informed features to enhance its predictive capability. Through a rigorous cross-validation process, we demonstrate the robustness and generalization of our approach. The framework's performance is evaluated using key metrics that assess both its ability to rank components by their failure risk and the accuracy of its probabilistic predictions over time. The results indicate that the developed digital twin possesses an exceptional ability to differentiate between high and low-risk components and provides highly reliable survival probability curves. This capability enables a significant paradigm shift from reactive to proactive maintenance, allowing for the optimization of asset management, the reduction of unplanned downtime, and a marked improvement in system safety. The proposed framework provides a compelling and generalizable solution for building effective digital twins that can guide decision-making in real-world engineering applications.