A fault-tolerant estimator is obtained from fusing the concept of fault detection with estimation. Two possible architectures are evaluated. At the center of the fault-tolerant estimation procedure is a bank of filters computing local state estimates, a residual screening scheme to isolate corrupted estimates and a method of blending untainted ones into a global minimum variance estimate, free of the effects of either sensor, plant or actuator fault. Previous voting schemes relied on statistical tests whereas the fault-tolerant estimation is based on a deterministic test on the local filter residuals. Finally, contrary to most fault-tolerant structures or multiple model approaches, estimating the fault is not necessary to ensure the quality of the global estimate.