The crack propagation of wing rivet holes has a significant impact on flight safety. However, accurate crack prediction is difficult. On the one hand, the calculated values based on traditional mechanics theory are conservative and have a large safety margin. On the other hand, the predicted values of deep learning do not match the mechanical theory, and the significant role of peak load in crack propagation has not been fully reflected. Therefore, this paper proposes a novel AK-TCN model for the crack prediction, considering the theoretical information constraints. Firstly, the “theoretical information assistance” technology is adopted. The calculated values (strength, life) based on mechanical theory and the measured values (stress) based on sensors are jointly used as inputs, enabling the AK-TCN model to fully extract valuable information. Secondly, the “multi-dimension feature fusion” technology is adopted. The theoretical information and measured information are mixed, and the importance of different information is adaptively adjusted, enabling the AK-TCN model to balance the foundational nature of theoretical information and the time-varying nature of measured information. Then, the “adaptive multi-scale convolution” technique is adopted. The temporal features in different periods are perceived by multi-scale convolution, and weights of them are assigned by attention mechanisms, so that the AK-TCN model can focus on certain key features that have a significant impact on crack propagation. Finally, the accuracy, effectiveness, and robustness of the proposed AK-TCN are verified by comparative experiments, ablation experiments, and noise experiments.
Read full abstract