Investigation of Time-series Prediction for Turbine Machinery Condition Monitoring

Abstract

Abstract A running compressor generates a huge amount of data every day. Some of the data contains useful information that allows the user to determine in advance whether the machine will fail or not. Forecasting in advance cannot only avoid the huge loss caused by unnecessary downtime but also ensure the safety of staff. The main tasks of compressor condition monitoring include data acquisition, data preprocessing, data analysis, alarm module, response strategy analysis, etc. Prediction is one of the tasks of condition monitoring. This paper mainly studies the performance of different time prediction models on the running data produced by a centrifugal compressor. In the data preprocessing part, this paper uses Bayesian wavelet denoising and probabilistic principal component analysis methods proposed by the research group to remove the useless parts in the data. In the part of time series model, the prediction performance of LSTM, GRU and ARIMA on the same dataset is compared. The comparative results indicate that since the dataset is non-stationary, the LSTM and GRU outperform the statistical ARIMA model.