Anomaly Prediction and Remaining Useful Life for Alternator Health Monitoring System

An alternator plays a vital role in conversion of mechanical energy into electrical energy in a generator. An effective model of Machine Health Monitoring System (MHMS) is paramount in providing insightful information for the plant operators as preventive measures against any failure. Several studies have shown the effectiveness of the data-driven model, over the physics-based model in the development of MHMS. This is due to the robustness and sensitivity of the data-driven model based on deep learning or machine learning in detecting the small changes of the data which is especially useful when there are changes in the working state of the machine. In this project, we will investigate two approaches in MHMS namely Remaining Useful Life (RUL) and anomaly prediction using deep learning techniques. Among the deep learning techniques, Convolutional Neural Network (CNN) and Long-Short Term Memory (LSTM) for RUL prediction and anomaly prediction are chosen to be further analysed. In particular, the performance of CNN-based MHMS will be presented in comparison with the LSTM-based MHMS. The proposed RUL model will be used to identify the total number of cycles left for the generator before a major breakdown. On the other hand, anomaly prediction based on LSTM would be developed based on the multivariate forecasting method. In short, CNN model works better for the RUL prediction with the lowest RMSE value of 41.66 and accuracy up to 99.89% as compared to LSTM model whereas for the anomaly prediction through multivariate forecasting, singlelayered LSTM has a better performance as compared to stacked LSTM with accuracy up to 94.54% and RMSE value as low as 0.113