In light of the problems of data monitoring and sensor fault diagnosis for thermal parameters in power plants, this paper builds an applied model based on mechanism analysis, kernel principal component analysis (KPCA), and radial basis function (RBF) neural network. First, auxiliary parameters related to the variable under study were obtained according to mechanism analysis. Then, KPCA was used to extract the high order nonlinear characteristics of the input variables, due to the high dimensionality, nonlinearity and strong coupling among them. Components were used to study and realize the reconstruction of thermal parameters through the RBF neural network. Last, sensor fault diagnosis was realized based on the prediction model and window moving method, and the fault data were able to be accurately replaced in time. Taking gas turbine outlet temperature as an example, the results show that this model performs with higher precision and generalization ability. Importantly, it can detect sensor faults and identify the type of fault in early stages, attaining a preferable detection effect.