Accurate prediction of fatigue life in materials is a fundamental challenge in mechanical design since it strongly affects the safety, reliability, and maintenance cost of engineering structures. Traditional empirical and analytical fatigue models, while widely applied, often fail to capture nonlinear and multiaxial effects arising under variable loading conditions. In this study, a hybrid deep learning architecture is developed that integrates three complementary components: a fully connected (FC) network for static material features, a bidirectional long short term memory (Bi LSTM) network for sequential loading path, and a transformer encoder for multi path feature fusion and high level relational learning. Experimental stress–strain data were normalized and divided into training and testing sets, and the model was optimized using the Adam algorithm with a learning rate of 5×10⁻⁴ for 2000 epochs. Quantitative evaluation demonstrates that the proposed FC–BiLSTM–Transformer model achieves superior accuracy compared with five baseline networks, with MSE = 0.335, MAE = 0.1385, and R² = 0.9470. Physically, the model captures complex fatigue responses without empirical hypotheses, enabling data driven representation of material behavior. The developed framework provides a reliable computational tool for fatigue life estimation and can be extended to complex materials and multiaxial loading conditions in aerospace and automotive applications.