To achieve effective detection of small faults in spacecraft attitude sensors and fault isolation of orientation sensors and inertial sensors in the presence of disturbance and measurement noise, we proposed a self-organizing recurrent neural network (SORNN) based small fault diagnosis method. Firstly, a SORNN model, including the self-organizing algorithm, termination condition, and adjustment condition, was designed to realize the adaptive adjustment of the number of hidden layer neurons and memory depth, thereby improving the fitting performance of the network. Then, a SORNN-based disturbance observer was designed for the kinematics subsystem. The network weight update algorithm was given, and the state estimation error convergence was proved. The output estimation error was passed through a low-pass filter to suppress the measurement noise of the star sensor. More rigorous residual and detection threshold were derived to improve the detection ability of small faults. Furthermore, a fault isolation observer was designed for the dynamic subsystem. The influence of unknown disturbance and noise on residual was eliminated by disturbance decoupling and disturbance observer compensation. The problem of fault isolation of different sensors was solved by using the redundancy relationship between dynamics and kinematics. Finally, the simulation results verified the effectiveness of the proposed method for detecting and isolating small faults of star sensors and gyros under the cover of disturbance and noise.