In response to the complex situation and low recognition efficiency of underwater robot thruster faults, this paper proposes a fault diagnosis method based on the Resnet18 network. This residual network can input the collected one-dimensional and multidimensional signals, and use a residual structure to greatly reduce the computational complexity of the network model. At the beginning of the fault, the fault status and category can be detected in a timely manner. After experimental analysis, This network model can achieve high accuracy in the “HaiZhe” AUV fault dataset in a short period of time.