Abstarct: Nowadays, safety and reliability are essential property in industrial equipment which is used in risky conditions such as aircrafts or automotive electrical systems, nuclear installations, ships, submarines. The purpose of fault diagnosis and tolerant control is to make ensure continuous system functionality, even after fault occurrence. in [2] the following definition is for fault: Unpermitted deviation of at least one characteristic property of the system. As the scale of system become larger, the number of faults growth, because of increasing control component such as sensors and actuators. So we should detect faults in order to cancel their effects or attenuate them until an acceptable level. Despite, the advancement of technology has led sensors to be less sensitive to external sources but measurement error can be cause due to sensor failure, broken or bad connections (especially in telecommunication applications) and some hardware or software malfunctions. All of these will be referred as sensor faults in this thesis. In the recent years, increasing concerns about sensor faults, lead some researchers to focus their efforts on developing sensor fault diagnosis and tolerant control in order that sensor signals are used in controlling the close loop system. Previous researches have divided fault detection into hardware and analytical redundancy methods. Methods such as fussy logic, parity equations and observers are related to analytical redundancy, whereas method such as wavelet transformation is related to other one. An easy way to fault detection and isolation is to compare process behavior with nominal process model and healthy conditions. Parity equation is one of the main methods in model baxse fault detection. in this thesis we propose a new hybrid fault diagnosis baxsed on structured parity equation’s approach and neural networks. An Elman recurrent network (ERN) evaluates parity equation residuals. The ERN trained offline baxsed on residuals generated from parity equations under both healthy and faulty conditions and estimating the fault size. ERN has certain unique dynamic characteristics over static neural networks, such as MLP, because of it’s context units. Structured parity equations cause better fault detection and isolation from primary parity equations. baxsed on simulation results we show the effectiveness of proposed method.