Failure prediction of auxiliary lube oil pump in gas turbines: Applying Monte Carlo neural networks to an unbalanced dataset
Prediction of future failure cases with associated time frames in order to adjust the maintenance plans is important for assets because the consequence of unplanned failure could be catastrophic and often result in large financial loss. Many industrial assets have multiple sensors for monitoring their condition, which allows for the collection of large amounts of data for failure analysis. However, there is a particular challenge for highly reliable assets as very few failure cases will be contained in their datasets. Such an unbalanced dataset makes it difficult to apply standard machine learning algorithms for failure prediction since such algorithms often depend on large number of failure cases for training the model parameters. This paper presents a solution to this problem using a Monte Carlo modified neural network, and applying this auxiliary lube oil pumps in industrial gas turbines. The results show that the model is able to predict the failures with an overall precision and recall values of 96.1% and 99.2% respectively.