Steam turbine is a crucial thermal rotating machinery in nuclear power plant. An abnormal long periodic vibration of steam turbine shaft occurred in nuclear power plants (NPP), which affects the economics and threatens the safety of the NPP. To address this issue, the characteristic of the abnormal periodic vibration is analyzed with empirical mode decomposition (EMD), and machine learning models based on long-short term memory (LSTM) neural network and feedforward neural network (FNN) are developed for fault diagnosis and attribution. The intrinsic mode functions (IMF) 3–6 obtained by EMD of the vibration signal reflect the characteristics of long-term abnormal vibration, while the relationship between six main operating parameters and the abnormal vibration is not reflected by EMD. Vibration data with 13,498 cases (24.9% abnormal vibration data) from nuclear power station are used to train and test LSTM and FNN models. FNN model well predicts the vibration state of the nuclear turbine with accuracy higher than 96.3%, while LSTM model has a low accuracy less than 82.4%. With the trained FNN, parameter sensitivity analysis is carried out and the results indicate that hydrogen pressure, oil pressure and oil-hydrogen differential pressure is vital for the abnormal long periodic vibration of the turbine.