The accurate detection of faults in robotic fish allows for improving the safety and reliability of its operations. This paper proposes a depth sensor fault diagnosis method based on Gramian Angular Field Fusion and Convolutional Neural Network (GAFF-CNN). Firstly, the depth sensor signals are augmented by a sliding window with overlapping data. Secondly, the one-dimensional time series sensor signals are converted into two-dimensional images by using Gramian Angular Field (GAF). To improve fault diagnosis accuracy and accelerate the training speed, using a weighted fusion method to fuse Gramian Angular Summation Field (GASF) and Gramian Angular Difference Field (GADF). After that, the model of CNN is established to train and test fused images for fault diagnosis. The result shows that the fault diagnosis accuracy is the highest at 97.22% when using a weighted coefficient of 0.3, and when the weighted coefficient is 0.4, the training speed is the fastest.