Decentralized state estimation of power systems is under investigation for the monitoring of modern power systems. In this framework, the identification of sensor failure is a critical issue. A novel method is proposed here to achieve this goal, yielding improved reliability of the decentralized power system monitoring. Particularly, the improved reliability could be regarded as the reliable uninterrupted state estimation of the power system, in the presence of failures of the sensors. Sensor measurements are locally validated before they are elaborated. The validation algorithm is based on reasonable thresholds of the measurand computed via polynomial chaos theory and determined based on the effect of the uncertainty in the system, particularly that of loads. These thresholds are dynamically computed; therefore, the criterion for acceptability of the measurements is always updated for the current operating conditions. In the presence of unacceptable data, the measurements are discarded and replaced with an estimate of the measured value. The application of this approach to decentralized state estimation is presented in this paper. Numerical results obtained from MATLAB/Simulink and real-time simulations of a shipboard direct-current zonal power system are used to demonstrate the effectiveness of the proposed method in addressing measurement failures in decentralized state estimation.