A system and method for predicting loose components in a nuclear power plant in advance are provided. The system comprises a first establishing module for establishing a neural network model, injecting normal vibration signal data, loose component signal data and simulated loose event signal data into the establishing neural network model for training to obtain a standard curve of a pre-loose vibration signal; a first signal acquisition module, configured to acquire vibration signals of components to be predicted in real time; a prediction module used for injecting the vibration signals of thecomponents collected in real time into the neural network model for identification and judgment, and predicting the possibility of the components loosening. The invention does not need a large amountof real fault data in the prediction process, obtains loose data of different stages of typical parts through limited real data and system mathematical model simulation, and identifies and predicts the typical parts by using a neural network method. It fills the gap that the monitoring of loose parts can only be judged afterwards and can not be predicted beforehand in this field.