Self-consistent outlier filtering machine learning approach for predicting functioning of water pumps

Contactinformatie

The application of machine learning in predictive maintenance opens the way for significant cost reduction and reducing down time in industrial and infrastructural processes. Hidden relations in monitoring data can be utilized for fault prediction and optimization of maintenance strategies. However, in reality it often occurs that datasets are partially corrupted by faulty data. Because of the complexity that these datasets often have it is not trivial to distinguish incorrect data from correct data. In this article a dataset that represents the functioning of water pumps is analysed and a model is trained to predict functioning. A self-consistent outlier filtering approach is developed to handle incorrect data while training the model. This approach is validated by showing that the correlation between predictors and the water pump functioning parameter is significantly higher after the incorrect data is identified and removed from the dataset. The developed self-consistent outlier filtering approach can be applied in principle for any data driven modelling problem.