Real-time fault diagnosis for autonomous underwater vehicles (AUVs) is crucial for ensuring overall system safety. As a critical component, the thruster operates is prone to failure under complex environments. The healthy status of underwater thrusters holds significant importance in enhancing the efficiency and safety reliability of AUVs. This study proposes a Physics-Informed Estimation Network (PIEN) to address thruster fault diagnosis. First, sensor data from the thruster are collected to establish corresponding fault mechanism models. The estimation network then constructs a predictive model to generate online residual data by physical parameters. Finally, a residual fault detector determines thruster malfunction status. The experimental validation with the "Haizhe" AUV dataset demonstrates PIEN's ability to quickly detect different thruster faults in prototypes. It also achieves better diagnostic performance than manually designed models.