A versatile data-driven model integrating domain knowledge and deep neural networks (DNNs) is proposed for fatigue life prediction with small samples. In the model, traditional fatigue life models, as comprehensive reflections of domain knowledge, are employed to generate pseudo labels for data augmentation. And a new DNN typology, called Branching neural network, is devised to distill useful training information without theoretical biases contamination. Moreover, further model improvement is achieved by the introduction of a subtractive clustering-based procedure for training data collection. The proposed model is experimentally validated in three case studies and shows better prediction performance against traditional models and conventional DNNs under small sample conditions.

提出了一种集成领域知识和深度神经网络 (DNN) 的通用数据驱动模型，用于小样本疲劳寿命预测。在模型中，传统的疲劳寿命模型作为领域知识的综合反映，被用来生成伪标签以进行数据扩充。并且设计了一种称为分支神经网络的新 DNN 类型学来提炼有用的训练信息，而不会受到理论偏差的污染。此外，通过引入基于减法聚类的训练数据收集程序，可以进一步改进模型。所提出的模型在三个案例研究中得到了实验验证，并且在小样本条件下显示出比传统模型和传统 DNN 更好的预测性能。