Enhancing remaining useful life predictions for nuclear reactor filters using knowledge models

Nuclear reactors worldwide are increasingly facing the challenges of aging infrastructure, with many in the latter stages of their operating lives. While modernization efforts have upgraded legacy systems to digital ones, there remains an opportunity to leverage this data for transitioning from time-based to condition-based, predictive asset management. Here, we explore the use of knowledge models to provide a structured framework for describing physical phenomena, enabling a reusable representation of the system state, and facilitating the application of traditional machine learning techniques. Current state of the art for remaining useful life (RUL) prediction on assets such as filters in nuclear reactors relies on custom development of software solutions, often utilizing curve fitting techniques, incurring high costs and significant time spent on data preprocessing. These solutions are effective but have limitations, such as hard-coded parameters, less accurate predictions early in an asset's life, and the need for manual data preprocessing. Additionally, these solutions typically rely on filtering out gaps in data and do not leverage potentially valuable information such as flow rate data. This work proposes an approach that utilizes knowledge graphs to model the relationships between data, concepts, and physical elements of the system, allowing for a modular and query-based solution. By capturing the temporal relationships in the graph, we enable automatic splitting of legacy data and identification of stable regions where predictions are unnecessary, forming a reusable description of outages, stable regions, and filter changes. In this particular case study, the focus is on improving RUL prediction methods for heavy water (D2O) filtersin a CANDU reactor by incorporating a knowledge model that captures descriptions of outages where filters do not degrade, thereby providing more accurate predictions of filter change dates. The proposed solution offers an avenue for explainability, as it provides engineers with a direct route from source data to model fitting. This work demonstrates the potential of knowledge models in providing a reusable and structured format for describing physical phenomena, enabling the application of traditional machine learning techniques and facilitating explainable decision support. Future work will examine the integration of flow rate data to account for variable filter degradation. The approach presented here contributes to the field of asset management in nuclear reactors and highlights the benefits of leveraging knowledge models for predictive maintenance.