A computationally feasible technique for the robust detection and estimation for actuator and sensor failures is presented. Model errors and component failures are represented by a bias vector called the failure state in system and measurement equations. A Kalman-Bucy filter is implemented to estimate the system state, and to generate the corresponding residuals. These residuals are then processed by using an adaptive fading Kalman filter to give the failure state estimate. The final state estimate is obtained by compensating model errors and component failures in the filter based on no failure assumption. The divergency of the filter based on no failure assumption is avoided by stepwise compensation of the failure state. The technique is applicable to the detection, estimation and compensation of slowly varying model errors and suddenly occurring component failures, and to the discrimination between them