As an important component of underwater vehicles, the fault diagnosis of underwater thrusters has been wildly studied. However, the overall diagnostic method for underwater thrusters considering sensor faults has largely remained unexplored. This paper proposes a new method, which first establishes the physical model of the underwater thruster and the load torque model. Based on this, simulating the initial deployment phase of the underwater vehicle, an adaptive Kalman filter (AKF) is used for thruster motor parameter identification to diagnose sensor faults. After confirming that the sensor’s health status is good, the underwater operating state of the vehicle is simulated. A variable structure sliding mode observer (VS-SMO) is used to decouple the vehicle’s speed and torque. At the same time, the load torque is observed online using the speed and current signals obtained from fault-free sensors. The recognition and classification of four fault modes are achieved through the variation of the load torque. Finally, the feasibility of the proposed aspects is verified through simulation experiments. The results show that the proposed method can effectively diagnose sensor bias faults, as well as the propeller entanglement, jamming, its breaking and falling off.