Marine air compressors work in a harsh environment, with multiple internal and external excitation sources. The collected noise signals are strongly time-varying, resulting in low fault diagnosis accuracy and thus difficulties in effective identification of various faults of marine air compressors. To address this problem, a fault diagnosis method combining Gramian Angular field (GAF) and Convolutional Neural Network (CNN) is proposed. Firstly, the basic principles, methods and implementation procedures of GAF and CNN are explained. Then, various faults of marine air compressors are simulated through experiments, and the noise signals of each fault are collected. The GAF is used to transform the one-dimensional time-domain signals into two-dimensional images, mapping the characteristic information to texture features such as colors and points of the two-dimensional images. Finally, the two-dimensional images are fed into CNN for feature extraction and fault diagnosis. The experimental results show that the proposed method can effectively identify various faults of the marine air compressor with a diagnostic accuracy of 99.2% superior to other algorithms, while ensuring the operational efficiency. It can provide technical references for the application of intelligent fault diagnosis systems for ships.