Abstract This paper presents an adaptive observer-based approach for the combined state estimation and active fault detection and isolation (FDI) of the individual-wheel-drive (IWD) vehicles. A 3-DOF vehicle model coupled with the Highway Safety Research Institute (HSRI) tire model is established and used as the observation model. Based on this model, the dual unscented Kalman filter (DUKF) technique is employed for the observer design to give fusion results of the interdependent state and parameter variables, which undergo nonlinear transformations, with the minimum square errors. Effectiveness of the proposed algorithm is examined and validated through co-simulation between MATLAB/Simulink and CarSim. The results demonstrate that the DUKF-based observer effectively filters the sensor signals, accurately obtains the longitudinal and lateral velocities, explicitly isolates the faulty wheel(s) and accurately estimates the actual torque(s) even with the presence of noise. Introduction The individual-wheel-drive (IWD) vehicle is an electric vehicle (EV) and has a powertrain that allows all the four wheels to receive torques exerted by the in-wheel/hub motors independent of the others. Recently, there are many studies investigating the torque distribution methods of this kind of vehicle, which is believed to have the potentials for handling and maneuverability improvement due to its dynamic redundancy [ 1]. Nonetheless, there is a strong possibility of motor failure with the increased number of actuators. An actuator fault makes the vehicle less stable and some faults may even result in fatal accidents. Thus, it is essential and of high priority to design the fault detection and isolation (FDI) system for IWD EVs, whereas it is also difficult to explicitly isolate the faulty wheel and estimate its torque because of the redundant configuration, too. By comparing the monitored system behavior and the estimated system behavior, a good FDI system can timely detect any abnormality and accurately locate and evaluate the faulty unit(s). Model-based FDI schemes employ the mathematical models of systems and they usually provide better results than data-based (model-free) schemes so long as the mathematical models are accurate enough. Generally speaking, there are three kinds of approaches used for the model-based FDI [ 2]: state observation [3 ], parameter estimation [ 4,5] and approaches based on analytical redundancy relations (ARR) or parity space equations [ 6,7]. In view of the vehicle controller design, effective operation of most active safety systems depends highly on an accurate knowledge of the vehicle states (namely the longitudinal and lateral velocities, and yaw rate). The kinematics method using the wheel speed signals is usually adopted for estimation of the longitudinal velocity of a conventional vehicle. However, if all wheels of an IWD EV are slipping or locked-up, which may likely happen, the velocity calculated by this method will be of less accuracy. Moreover, it is infeasible to acquire the lateral velocity, which is crucial for the lateral dynamics control, by means of a kinematics method. Not surprisingly, the longitudinal and lateral velocities can be directly measured by some sensors, such as the optical speedometer and the Global Positioning System (GPS). But their applications in real-time control systems are severely limited either by the high cost and/or by the unreliability. Thus, they are still impractical for mass production. As a result, many observer- based estimation methods have been put forward to obtain vehicle states [8]. Among them, Kalman filter (KF) has gained considerable attention because it is good at dealing with both measurement noise and system uncertainties [ 9]. Meanwhile, extensions of the KF to nonlinear systems have been sought; in particular, the unscented Kalman filter (UKF) was proved to obtain better filter performance in terms of both accurac