The turbine driven boiler feed water pump set is an important auxiliary part of the thermodynamic cycle system in thermal power plants. Effective fault prediction for the turbine driven boiler feed water pump set is helpful to its condition-based maintenance. In this paper

by a feature extraction method based on statistics and a feature selection algorithm named Relief

the history data in plant-level supervision information system(SIS) is converted to the input parameters of classification models rationally. Moreover

combing with the actual fault cases in two power plants

the blades fracture of small turbine in turbine driven feedwater pump set and dynamic/static balance disk rubbing in feedwater pump

five classification algorithms are applied to establish the classification models for distinguishing normal state and fault state. The actual data verification shows that

the BP neural network

support vector machine(SVM) and combined classification algorithm have better performance than other methods

which can identify the potential risks of faults 4～10 weeks in advance. The result provides new idea for other equipment failure prediction.