Active steering fault diagnosis via integrated LSTM-based sensor detection and robust actuator fault estimation

Active Steering Systems are essential to the advancement of automated vehicles, as they directly influence vehicle dynamics and safety by enabling precise control of the vehicle’s trajectory. However, faults in sensors and actuators can affect performance and safety. Previous studies have primarily focused on methodologies for detecting or estimating actuator faults. Nevertheless, sensor faults are often overlooked even though they can compromise data reliability and lead to incorrect fault diagnosis. Due to this reason, this paper presents a new fault diagnosis methodology that combines a robust switched Luenberger observer with an unknown input observer to estimate vehicle states and actuator faults. To prevent false estimations due to sensor faults, an LSTM neural network is used to detect sensor faults. These detections allow the system to switch the configuration of the observer, thereby preventing incorrect actuator fault estimates. The methodology is validated in a Hardware-in-the-Loop test using the vehicle dynamics software CarSim®.