Abstract: At present, nuclear power plant equipment sensors generate a large amount of monitoring data every day, but the utilization of these data is relatively low. Research on using monitoring data for fault prediction of important equipment in nuclear power plants is still in the exploratory stage. In response to this situation, this article takes the main feedwater pump of nuclear power plants as the research object, preprocesses the operation status data, reduces the various monitoring data dimensions, and then predicts whether the equipment will malfunction in the future through multiple selected machine learning models. The selected models are Linear model, support vector machine (SVM) model, and LSTM model. For the fault diagnosis of the main feedwater pump, the discrimination criterion is that the vibration signal of the feedwater pump bearing exceeds the threshold. When the predicted results of the model exceed the threshold, an early warning message is issued to remind the operation and maintenance personnel of the nuclear power plant to strengthen their attention, take timely fault diagnosis and maintenance measures, and prevent serious consequences caused by sudden equipment failure and shutdown, which will affect the safe operation of nuclear power plants. Specifically, this study utilized the operating parameters of the main feedwater pump during operation, which are more than ten parameters other than the vibration signals of the feedwater pump bearings. Through a predictive model, the vibration situation of the feedwater pump bearings was predicted, and the fault status of the main feedwater pump was determined based on the vibration prediction signals to evaluate whether the equipment can operate for a long time. After preprocessing and feature selection, the selected dataset was trained and validated against traditional data processing models and currently popular deep learning models. After evaluation, it is found that the deep learning model LSTM has more accurate prediction results, with mean square error and goodness of fit meeting the required range and could better fit the actual operation of the equipment. At the same time, by using the ridge regression algorithm, the parameters that have a significant impact on the normal operation of the main feedwater pump can be analyzed. In the actual work of the equipment, it can help operation and maintenance personnel find the cause of faults and improve the quality of equipment operation and maintenance. The historical data used in this study mainly cover common typical faults of the main feedwater pump. In the future, it is necessary to continue collecting other types of fault data for model training and validation, so that the model can cover all possible faults that may occur in the main feedwater pump. Overall, the research prospects for fault prediction of such large-scale equipment are very broad.