Orthotropic Steel Decks (OSDs) are highly prone to fatigue cracks at the deck-rib connection (RD) during service, significantly compromising structural safety and durability. To overcome the limitations of traditional fatigue life assessment methods based on S-N curves and nominal stress approaches for this critical detail, this study proposes a data-driven method for evaluating the crack fatigue life of RD connections. By integrating a Bidirectional Gated Recurrent Unit (BiGRU) and a Self-Attention mechanism (SA) into a Convolutional Neural Network (CNN), and leveraging a dataset constructed from simulation experiments, the proposed approach aims to predict the fatigue crack life of RD connections. The specific research process is as follows: First, a finite element model of the RD connection is established using ABAQUS software, and fatigue crack propagation analysis is conducted using FRANC3D to generate a relevant dataset. Second, a CNN-BiGRU-SA model is constructed, where the CNN extracts local features from the dataset, the BiGRU captures temporal dependencies in crack propagation, and the SA suppresses irrelevant information while ghlighting critical features, thereby obtaining more salient feature representations. Finally, the trained CNN-BiGRU-SA data-driven model is employed to predict the fatigue crack life of RD connections. The results show that the average error on the validation set is approximately 350,000 cycles, with the maximum error controlled within 1.24 million cycles. On an independent test set, the error converges to around 20,000 cycles, with the peak error not exceeding 46,000 cycles. Ablation experiments confirm that the synergistic effect of BiGRU and SA significantly enhances prediction accuracy. This study provides an efficient and precise data-driven solution for evaluating the fatigue crack life of critical details in orthotropic steel bridge decks.