The safety of nuclear power plant operation can be improved efficiently by developing the fault diagnosis applications. However, adopting single technique may restrict the flexibility and accuracy of the fault diagnosis module. This paper proposes a new fault diagnosis methodology for nuclear power plants based on Kernel Principle Component Analysis (KPCA) and Quadratic Support Vector Machine (QSVM). The nonlinearity between variables is analyzed and the importance of variables are quantified by standard deviations. Sensor faults are distinguished from system faults using contribution charts. In pre-process stage of classification, five normalization modes are compared to achieve better performance of classifiers. For fault identification, samples are transformed to QSVM classifier and the diagnosis results are compared with other classifiers. To further evaluate the performance, three typical case conditions are tested. QSVM outperforms other classifiers on all types of system faults. The findings are beneficial for fault diagnosis of nuclear power plant.