This thesis deals with sorne methods for improve reliability in electrotechnical devices. The proposed methods are based on analytical redondancy. An approach using astate estimation is developped in order to detect failures on a statie converter-fed DC motor on wich we considered sensor faults and power electronic failures. Chapter II deal with parallel state model and propose two sort of these : "global" and "decoupled~~ model. The second one allowed a good fault detection and localization but was disturbed by load torque variations. Robustness of this model with respect to parameter variations is discussed and results of simulation are shown. Chapter III present an approach to be robust against disturbances : using an Unknown Input Observer. UIFDO allowed good results in term of disturbance decoupling with respect of a good fault detection. This work conclue on the capability of analytical redondancy to detect and localize faults on electric drives
