With the wide application of fibre-reinforced composites in aerospace, the fatigue problem of composites is becoming more and more prominent. In order to achieve efficient and accurate fatigue damage analysis, a data-mechanism driven method for the progressive analysis of fatigue damage in composites is proposed. The method utilizes a single-hidden-layer neural network as its fatigue constitutive law for simulations of fatigue delamination under cyclic loading. In order to achieve neural network training with a small quantity of samples, this paper uses a Paris-law-informed regulations to achieve data-mechanism fusion for neural network model training. The ability to analyze fatigue delamination is validated in in the full range of mode-I and mode-II as well as mixed modes of different mode ratios using double cantilever beam (DCB) and 4-point end flexure (4ENF), the paper further verifies the applicability of the cohesive model in the case of complex fatigue delamination front using an reinforced double cantilever beam (R-DCB) model. The numerical results of this paper show that the data-mechanism driven fatigue damage progressive analysis method for composites could rapidly and effectively simulate the composite delamination propagation with high fidelity, providing a new idea and method for composite structure design and safety assurance.