A unified prediction approach of fatigue life suitable for diversified engineering materials

Abstract

Investigation on fatigue life prediction approaches for welded joints of different materials was of great significance for the reliability and safety of engineering structures. A fatigue performance dataset with sufficient characteristics was established via data processing and augmentation. The modification of loss function and the physics-informed influencing factors was used to realize the multi-level physical intervention on the prediction model. The importance of physics-informed influencing factors under mechanical-properties dataset was analyzed and integrated into the deep convolutional neural network (DCNN) framework via attention model. The significance of the intervention using physics-informed influencing factors and the identification using mechanical properties was proved via the comparison of different prediction methods. Further, a mechanical properties-based prediction method for fatigue life of multiple materials was established, and the average prediction error (30.5%) and standard deviation of prediction error (16.1%) of the proposed approach were significantly lower than that of the other prediction methods. By the identification via mechanical properties and the introduction of physics-informed influencing factors, the machine learning method can be used to evaluate the fatigue life of multiple materials in engineering with low consumption.