Abstract: An improved approach to sensor fault detection and isolation based on particle swarm optimization/least squares support vector machine (PSO-LSSVM) is presented to solve the problem that the output of sensors is volatilized with the productive process, resulting in false high rates of sensor fault. First, principal component analysis (PCA) is used to model sensor output and detect sensor fault in the case of noise. To reduce the failure rate of fault isolation, PSO-LSSVM is adopted to predict the output sequence. The difference between predicted values and true values acts as residual vector, and then fault mapping method is used for fault isolation. In the simulation, the method is used to detect and isolate output data of different sensor separately, based on actual production data collected from an oil extraction platform of a Liao River oil field. Results show that this method can detect the fault effectively and improve the accuracy of isolation.