In the general stochastic nonlinear and non-Gaussian systems, the sensor faults including biased and scaled readings caused by sudden calibration errors have adverse effect on the precise monitor and stable control of the system. To deal with this problem, a novel diagnosis and identification method is proposed. An adaptive particle filter is developed to calculate the difference between the measurements and the particle filter estimates, then the type and magnitude of sensor faults are determined through with maximum likelihood estimation, thus the fast and precise detection of sensor fault is realized, and the adverse effect caused by the faults can be compensated. Some simulations are carried out on a boiler model, and the results validate the effectiveness of the proposed method.