The vibration trend prediction of hydropower units based on wavelet threshold denoising and bi-directional long short-term memory network

Abstract

The faults of hydropower units are usually reflected in the form of vibration. Therefore, the accurate prediction of vibration trend can improve the recognition of the operation state of hydropower unites, which provide strong support for condition-based maintenance. In this paper, a framework based on a fused metric for feature selection and the bi-directional long short-term memory network (BLSTM) is developed to obtain the multi-step trend prediction for hydropower units. Initially, the wavelet threshold denoising (WTD) method is used to eliminate the interference of strong background noise from the raw sensor data. Next, a fused metric based on Pearson and distance correlation coefficient is proposed to pick out suitable working condition variables to make the prediction model more stable. Ultimately, the BLSTM network is used to predict the trend of vibration. In order to evaluate the prediction performance of the model, the vibration monitoring data of a pumped storage hydropower station are collected for comparative experiments, which proves that the proposed method has good prediction ability and generalization ability, which is suitable for the trend prediction of vibration of hydropower units.