The high-precision path tracking control of unmanned rollers relies on the accurate measurement of the vehicle's location information by sensors. However, in the harsh plateau dam environment, the positioning sensors' signals are prone to sudden changes and loss, which poses a great threat to the safety and stability of the system. For this reason, a fault-tolerant control method based on signal compensation by using self-learning model was proposed. First, according to the drift characteristics of the steering system reflected in the operation of the unmanned roller, the complex hydraulic steering system was simplified, and a linear steering model with flow loss parameter was constructed. Then, an online learning algorithm was proposed for mainly learning the flow loss characteristics of the system. Finally, a cascaded disturbance rejection controller was used to realize the path tracking control. Both simulation and real vehicle experiments were carried out. The results show that the vehicle can continue high-precision running within 40 s after the sensor fails. Compared with the cases of no fault-tolerant control and fault-tolerant control without model self-learning, the running time is extended by 18.7 times and 2.7 times respectively, which greatly improves the safety and efficiency of the system.