Abstract: Compared with the past, there are more important parameters available for collection and monitoring during the operation of nuclear power plant due to the development of sensor technology. This situation not only increases the task of the operator, but also increases the load on the monitoring system. Since most parameters are correlated and some are redundant, the effective information in the parameters can be represented by a few parameters. In response to the above premise, the sparse autoencoder was used in this paper to extract the features of operating parameters of nuclear power plant. These feature data were then used in condition monitoring. The results show that using the data obtained by feature extraction for condition monitoring can not only improve the accuracy of state monitoring, but also reduce the computing resource, and this conclusion is applicable to both single and multiple normal working conditions. The results have important guiding significance for improving the safety of nuclear power plant.