A new method proposed here detects, reconstructs, and identifies faulty sensors using a normal process model, which can be built from first principles or statistical methods such as partial least squares or principal component analysis. The model residual is used to detect sensor faults that demonstrate a deviation from the normal process model. To identify which sensor is faulty, a structured residual approach with maximized sensitivity is proposed to make one residual insensitive to one subset of faults but most sensitive to other faults. The structured residuals are subject to exponentially weighted moving average filtering to reduce the effect of noise and dynamic transients. The confidence limits for these filtered structured residuals are determined using statistical inferential techniques. In addition, other indices including generalized likelihood ratio test, cumulative sum, and cumulative variance of the structured residuals are compared to identify faulty sensors. The fault magnitude is then estimated based on the model and faulty data. Four types of sensor faults, including bias, precision degradation, drifting and complete failure, are simulated to test this method. Data from an industrial boiler process are used to test its effectiveness. Both single faults and simultaneous double faults are detected and uniquely identified with the method.