In recent years, intelligent navigation security is a hot spot in the field of water safety protection. Distributed optical fiber acoustic sensing technology based on phase sensitive optical time domain reflectometer can realize distributed monitoring of multi-point disturbance along optical fiber. However, due to the complex water environment and the fading of system signals, it is difficult to identify the disturbance signals stably and effectively. The main noise sources of distributed optical fiber acoustic sensing system are interference fading and polarization fading. Both of these phenomena greatly weaken the signal at the fading point, resulting in signal distortion. Under water, the cable is less coupled with the environment, and is more susceptible to the influence of water waves, currents and other factors. Therefore, timely and effective elimination of fading interference plays an important role in the identification of water security events. Combining fading mask, attention mechanism and self-supervised learning, this paper proposes a distributed optical fiber acoustic sensing system based on fading mask autoencoder for ship security event recognition in waters. This method is aimed at the ship event signal in the water area, and generates a basically noise-free signal by shielding fading noise into the deep learning model, so that the model can learn directly and greatly reduce the influence of fading noise on signal recognition. Mask autoencoder is an extensible self-supervised learner. It combines the attention mechanism in the form of a mask to achieve high-precision training with the simplest coding-decoding model structure, while it can also transfer learning only through the model weights of the encoder. On this basis, the fading mask autoencoder method is more helpful for the distributed optical fiber acoustic sensing system to achieve effective event recognition. Firstly, the amplitude signal of distributed optical fiber acoustic sensing is analyzed to determine the fading position. The model is then pre-trained using the upstream task of the fading mask autoencoder. The basic characteristics of distributed optical fiber acoustic sensing signal are learned by means of random mask. Finally, the downstream task of fading mask autoencoder completes the intelligent event recognition training by fading position mask. In this paper, four kinds of ship security events collected at the water test site are used as classification data, and the fading mask autoencoder is compared with the mask autoencoder and three related models. The results show that the average training accuracy of the fading mask autoencoder is 98.34%, which is 4.9% higher than that of the mask autoencoder. The average training loss was 0.1094, which was 0.095 less than that of the mask autoencoder. The average test accuracy was 93.01%, which was 6.45% higher than that of the mask autoencoder. Compared with the other three models, the fading mask autoencoder has higher training accuracy, lower Loss and faster convergence speed. Its average performance index is about 4.88%-7.62% higher than other models. Therefore, the fading mask autoencoder model based on the improved mask strategy can extract useful information from the signal more efficiently and accurately for training, and has better stability and generalization, which is suitable for the identification of navigation security events in waters.