On the use of data-driven machine learning for remaining life estimation of metallic materials based on Ye-Wang damage theory

Abstract Remaining fatigue life estimations are crucial for guaranteeing the operation reliability and structural integrity of engineering components. In this respect, physical process-based damage theories and data-driven machine learning (ML) techniques are two useful tools for damage accumulation analyses. However, conventional damage theories are generally limited to specific materials and loading conditions due to over-restrictive assumptions and laborious calibration procedures; on the other hand, ML techniques are easy to be afflicted by the inadequacy of training data in practice. To overcome these limitations, a damage theory-informed machine learning (DT-iML) model is proposed in this work, which enables the integrated utilization of domain knowledge and raw data information. In the model, a damage theory, i.e., the Ye-Wang theory, is introduced as the baseline for optimizing the data-driven processes of model training and prediction. Moreover, the subtractive clustering is further integrated to stabilize the model performance through constructing a global-representative training dataset. Extensive experimental results of nine metallic materials under two-step loading are collected from the literature for model evaluation and comparison. The results demonstrated that the DT-iML model allows for lower data requirement and higher performance stability than purely data-driven ML model, and meanwhile show better estimation accuracy and robustness than conventional damage theory.