The reliability and safety of active steering systems are fundamental to the successful operation of modern automated vehicles, since these systems directly control vehicle dynamics and trajectory. However, the increasing complexity of these systems increases the likelihood of faults occurring in the actuators, which can severely impact performance and compromise vehicle safety. Therefore, it is essential to accurately detect and identify these faults in order to enable effective fault-tolerant control strategies. Furthermore, sensor anomalies pose a significant threat, as unreliable sensor data can lead to incorrect estimations of actuator faults and compromise diagnostic integrity. This paper presents a novel, integrated fault diagnosis architecture that addresses these two challenges. The framework’s primary objective is robust actuator fault estimation using a dual-observer structure that combines state and fault estimation. This is synthesised using linear matrix inequalities. To prevent misdiagnosis caused by sensor errors, the framework incorporates a long short-term memory neural network dedicated to real-time sensor fault detection. When an anomaly is detected, the neural network output activates a configuration switch in the main observer, effectively isolating the unreliable sensor data. This integrated approach ensures that the diagnostic system remains highly accurate and robust. The efficacy of this complex methodology has been confirmed through extensive hardware-in-the-loop validation using CarSim, which is high-fidelity vehicle dynamics simulation software.