Long Short-Term Memory Networks for Facility Infrastructure Failure and Remaining Useful Life Prediction

Abstract

Sensors attached to an asset acquiring vibration patterns during both operational and failure states have been used to diagnose fault conditions and to predict future failures of the components being monitored. In this research, we investigate Long Short-Term Memory (LSTM) networks, a type of Recurrent Neural Network (RNN), for failure diagnosis and remaining useful life (RUL) prognosis of such deteriorating components. LSTM networks’ long-term dependency capability, which allows LSTM’s to recall information for long term sequence lengths, can also be used to predict the probability of failure within a specified time frame. In this paper, we develop and apply a stylized LSTM model to a motor degradation dataset for the purposes of diagnosing failure and predicting the probability of failure within a specified time frame, as well as predicting RUL. We developed the dataset by acquiring automated sensor measurements from an induction motor attached to a destructive test platform. The performance of the LSTM model on the developed dataset is compared to that of the Random Forest (RF) algorithm as RF is reputably known for classification and regression. The results demonstrate that the LSTM provides quality predictions of motor failure, failure probability and RUL on the developed dataset. When compared to the RF approach, the LSTM performs comparably well in failure classification and outperforms the RF in RUL prediction.

Keywords

• Predictive maintenance
• deep learning
• long short-term memory
• condition monitoring
• rotor bar failure
• Industrial Internet of Things