This paper proposes a fault diagnosis method based on regularized threshold wavelet decomposition, Markov transition field (MTF), and SE-DenseNet model, to address the problem of low diagnostic accuracy of ship power gearboxes caused by variable noise and complex working conditions in their operating environment. First, the squeeze-and-excitation networks (SENet) module is integrated into densely connected convolutional networks (DenseNet) to improve the model's feature extraction capability for MTF images. Regularized threshold wavelet decomposition is then used to denoise the original data. The preprocessed one-dimensional time-series data is transformed into a two-dimensional MTF image, which is fed into the improved DenseNet model for fault diagnosis, with a classification output head achieving accurate identification of fault types. The model is experimentally validated using the gear fault dataset from Southeast University and the fault data collected by a self-built ship power gearbox fault analysis unit. On the Southeast University dataset, the accuracy reaches 98.8% and 96.6% under two different working conditions, respectively. On the self-built dataset, the accuracy reaches 97.43% and 98.14% under two working conditions, respectively. The results show that the proposed model can identify fault types with high accuracy. Compared with other algorithm models, it exhibits good accuracy and generalization under different working conditions, providing a reference for the fault diagnosis of ship power gearboxes.