Optimal Input Selection for Recurrent Neural Network in Predictive Maintenance

Abstract

Rotating equipment are essential components in oil and gas industry processes. Failure of rotating equipment can lead to critical downtime of a plant. Traditional preventive and reactive maintenance practices are ineffective in identifying equipment faults. On the other hand, predictive maintenance receives highly attention in the industry because it learns from the past information to predict future failure and subsequently early maintenance can be performed before the occurrence of equipment failure. In this paper, 14 parameters from air booster compressor (ABC) are retrieved for optimal input analysis and selection. A sequential of data are recorded in a real-time manner for prediction. However, training a data model with high dimensions requires vast time and complexity. Therefore, optimal input variable selection for recurrent neural network (RNN) model is performed using principal component analysis (PCA) for signal prediction of ABC motor parameters. The performance of the model is compared with large set of inputs were used. The obtained results showed that the selected input variables based on PCA results in lower error rate and better prediction accuracy.