Predictive maintenance of standalone steel industrial components powered by a dynamic reliability digital twin model with artificial intelligence

The increasing use of Artificial Intelligence algorithms underscores the importance of large datasets. Recent trends highlight the need for high-quality training data, especially in scenarios where data may be outdated or insufficient. This challenge is particularly evident in applications where sensors cannot be installed or data is limited, such as in the case of steel components widely used in various industries. To address this gap, modelbased approaches show promise by using advanced Digital Twin systems to generate synthetic data, closer to the real working scenarios, for training Artificial Intelligence algorithms. This paper introduces a novel Dynamic Reliability Digital Twin to model cumulative fatigue damage in steel components based on Wo center dot hler and Manson & Halford theory and on a Monte Carlo simulation, providing a dataset for training an AI predictor to estimate remaining useful life. The results demonstrate that machine learning algorithms yield favorable outcomes when the dataset is appropriately calibrated. Therefore, a thorough understanding of the underlying physics is essential to avoid potential bias in the machine learning results.