A novel fretting fatigue life prediction neural network based on data extended agent model and intense physical constraint

Fretting fatigue is a critical failure mode in mechanically joined structures that severely affects component reliability and safety, making life prediction essential. Due to its highly complex damage mechanisms, traditional prediction methods are limited in accuracy. Although machine learning has recently been applied to fatigue prediction, its performance strongly depends on large-scale, high-quality datasets. However, available fretting fatigue datasets are still small, which significantly restricts the application of data-driven approaches. To address the challenges posed by small-sample limitations, this study employs aluminum alloy fretting fatigue life prediction as a representative case. First, a stress-driven Artificial Neural Network (Stress-ANN) surrogate model is proposed to establish mapping relationships between fretting loading parameters and stresses within the fretting contact zone. Leveraging the aluminum alloy’s fretting fatigue S - N curve, data samples are augmented via a stochastic multiplier method. Subsequent data cleansing optimizes dataset quality, achieving effective fretting fatigue dataset expansion. Second, a Physics-Informed Neural Network (PINN) architecture tailored for fretting fatigue life prediction is developed. A strongly physics-informed loss function is constructed, incorporating both a physical loss term ( LOSS _phy ) and a physical boundary loss term ( LOSS _bou ). Results demonstrate that the augmented dataset significantly enhances model prediction accuracy and generalization capability. Furthermore, the introduction of the physical boundary constraint loss term facilitates superior model performance. In summary, the proposed methodology integrating dataset augmentation with physics-informed constraints effectively improves the efficiency and accuracy of fretting fatigue life prediction, laying a robust groundwork for subsequent in-depth research.