A CNN-BiGRU-MHSA Hybrid Model-Based Method for Fatigue Life Prediction of Different Structural Components

Under the effects of vibration, loading, and other stimuli, structural components are susceptible to surface crack initiation and progressive propagation, resulting in reduced service life. To accurately predict fatigue failure behavior, a fatigue life prediction approach for cracked structural components is presented, which combines Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), and Multihead Self-Attention (MHSA). Firstly, multi-source data undergoes feature extraction via convolution operations, with the activation intensities of visualized features utilized to conduct characteristic mapping analysis corresponding to the fatigue tensile physical process. Secondly, the BiGRU method, which can simultaneously capture forward and reverse information flows, is adopted, and the MHSA mechanism is introduced. Specifically, input data are mapped to multiple differentiated feature subspaces through distinct linear transformation matrices, enhancing the sensitivity of key samples within the global features. Subsequently, the proposed method is validated using the aluminum alloy 2024-T3 thin plate dataset, the steel wire fatigue test dataset, and a self-constructed fatigue tensile test dataset. Meanwhile, its prediction accuracy is compared against other machine learning methods, and the results demonstrate that the proposed method exhibits superior prediction accuracy and robust performance. Finally, scanning electron microscopy (SEM) is employed to observe the microtopography of the material's fracture surface. Based on the analysis, the intrinsic relationship between microscopic fatigue fracture mechanisms and the performance of macroscopic prediction models is revealed.