Prediction of fatigue crack damage using in-situ scanning electron microscopy and machine learning

Abstract

Nickel-based single crystal superalloys, as engine blade materials, are prone to fatigue damage due to repeated startups and shutdowns. Therefore, monitoring and quantitatively estimating fatigue cracks are essential for engineering structures to ensure safety. In this study, we proposed a method for fatigue crack segmentation and damage prediction based on deep learning and in-situ high-temperature scanning electron microscopy (SEM). Sequential SEM images describing the crack initiation and propagation under near-service conditions were obtained by conducting in-situ high-temperature fatigue experiments. A fatigue crack dataset with high-quality was thus constructed for further dynamic and real-time crack segmentation and damage assessment. Deep learning-based models were used to segment cracks and predict damage behavior (i.e., crack area, length, width, and stress intensity factors) at future based on prior damage information. The short-term and long-term damage prediction capability were validated by comparing model performance when predicting damage at different future time points. Additionally, we compared the model performance when predicting damage at specific time point based on varying lengths of input sequence. Results demonstrated that the model could segment cracks and scales with different sizes accurately. The model performed well in short-term damage prediction. The long-term predictive performance showed decrease than that of short-term, which could be improved by feeding long length of input sequence. The proposed approach demonstrates the feasibility and effectiveness of deep learning-based crack segmentation and damage prediction, which facilitates the move toward real-time analysis and rapid diagnosis of material damage in the future.