Abstract:In power plants, abnormal bearing vibrations often occur when a turbine malfunctions. To take preventive measures before such malfunctions happen, accurately predicting turbine bearing vibrations becomes the foundation for fault prevention. However, the significant fluctuation in vibration data makes precise prediction challenging. To address this issue, the approach involves initially decomposing the vibration signal into multiple components using Empirical Mode Decomposition (EMD). Subsequently, these components are analyzed using Sample Entropy (SE), and the resulting components are reconstructed into trend and fluctuation signals. Next, employing Long Short-Term Memory networks with an attention mechanism (LSTM-Attention), the vibration-related measurement data of bearings are used as auxiliary variables to individually forecast the trend and fluctuation signals. Finally, the forecasted trend and fluctuation signals are combined to obtain the predicted vibration data. To validate the feasibility and accuracy of this model, experiments were conducted using real operational data from a power plant in Northwest China. The experimental results demonstrate that the proposed model exhibits smaller errors compared to traditional models.