A Data-driven Approach to Predict Maintenance Delays for Time-based Maintenance

Nuclear power plants ensure safety and reliability through Time-Based Maintenance where maintenance activities are carried out in a recurring scheduled manner. However, with aging reactors and the resource, financial and safety risks associated, maintenance items are often delayed which has subsequent issues to system reliability. This work explores the use of Machine Learning algorithms on a representative dataset that have similar data types to that of nuclear maintenance data. The results of the prediction models show that Deep Neural Networks and Random Forest Regression models provide a low prediction error (Mean Average Error). With the results of the prediction models, it was determined that the use of machine learning should be explored further with real maintenance data as the computational costs are relatively low. In addition, a framework was developed on how to implement and use prediction models for improving time-based maintenance schedules. This work acts as a conceptual foundation to introduce machine learning tools for improving maintenance planning and decision making.