Fatigue crack diagnosis (FCD) is of great significance for ensuring safe operation, prolonging service time and reducing maintenance cost in aircrafts and many other safety-critical systems. As a promising method, the guided wave (GW)-based structural health monitoring method has been widely investigated for FCD. However, reliable FCD still meets challenges, because uncertainties in real engineering applications usually cause serious change both to the crack propagation itself and GW monitoring signals. As one of deep learning methods, convolutional neural network (CNN) owns the ability of fusing a large amount of data, extracting high-level feature expressions related to classification, which provides a potential new technology to be applied in the GW-structural health monitoring method for crack evaluation. To address the influence of dispersion on reliable FCD, in this paper, a GW-CNN based FCD method is proposed. In this method, multiple damage indexes (DIs) from multiple GW exciting-acquisition channels are extracted. A CNN is designed and trained to further extract high-level features from the multiple DIs and implement feature fusion for crack evaluation. Fatigue tests on a typical kind of aircraft structure are performed to validate the proposed method. The results show that the proposed method can effectively reduce the influence of uncertainties on FCD, which is promising for real engineering applications.