The underwater thruster is an essential driving element for underwater platforms. Since underwater thrusters may fail because of external factors, a fault detection system is necessary for reliability and safety. Among the underwater thruster fault detection and diagnosis methods, a data-driven learning method, which does not require expertise or a physical model of the platform, is applied because a rule-based method lacks flexibility and a model-based method relies heavily on expertise. Although high-quality, large-capacity datasets are essential to implementing data-driven fault detection systems, the amount of fault sensor data is relatively scarce because most underwater thrusters operate in a normal state. However, if the platform is operated in a fault state for a long time to acquire fault sensor data, performance degradation of the thruster or accidents may result. In this study, we investigated a fault detection system wherein a small number of vibration sensor datasets were used as inputs for a generative adversarial network (GAN), and new vibration sensor datasets were generated, extended, and applied to a long short-term memory neural network for fault detection in an underwater thruster. For the defects detected by the machine learning algorithm, the rotor imbalance due to a thruster blade fault or the entanglement of floating objects was analyzed. To collect the vibration sensor dataset of the thruster, a structure for an underwater experiment was designed, and a system with a stable power supply, thruster control, and the capability to acquire vibration data was developed. Vibration sensor data obtained from the experiment and those generated by the GAN were comparatively analyzed in terms of their vibration characteristics using the fast Fourier transform. After training the neural network with GAN-generated data, the fault detection system was validated using real data as prediction data.