Nuclear plant operators require trusted data analytics tools to support the management of asset health throughout their operating lifetimes. Management of the data pipeline that serves data analytic tools, alongside the development of the analytic tools themselves, creates an ecosystem whereby operators can more effectively access the risk associated with the utilisation of data-driven systems within their decision-making processes. Prognostics and health management, and structural health monitoring practices allow nuclear power plant operators to monitor the state of assets and structures in the plant to avoid the financial strain and loss of generation from unexpected faults. However, for these technologies to be adopted, they must have high accuracy to prevent false alarms or missed faults, which can degrade operator trust in these tools. There is a need for trustworthy analytics across the nuclear sector, with analytic tools capable of uncertainty quantification to attribute risk to analytic outputs, and an understanding of uncertainty sources in the full data pipeline serving these analytics. This work firstly investigates the impact of data pipeline design on analytic performance by using a SHAP-based explainability tool to form part of a novel pipeline design interrogation framework. This framework identifies the highest impact positive and negative performance drivers, providing informed design decisions for data pipelines to improve performance of analytic tools within these pipelines. The process was shown to be transferable to the data pipeline designs of similar assets with less available design data, leveraging insights from one system to reduce the uncertainty sources within designs across other systems for improved fleetwide monitoring. Secondly, this work demonstrates the development of a novel copula-based calibration module within a hierarchical modelling structure which is used to improve predictions of transparent, but interchangeable, base models that are commonly applied within the highly regulated nuclear sector. The approach has the additional benefit of uncertainty quantification which attributes risk to the final prediction. The procedure was shown to be effective for spatial and temporal data, demonstrating applicability to a diverse set of engineering applications. The methods developed in this work have made progress towards providing trustworthy data analytic tools and data pipeline designs to provide nuclear operators with the risk associated with applying such tools to the management of the health and maintenance of their assets.
| Date of Award | 9 Jun 2025 |
|---|
| Original language | English |
|---|
| Awarding Institution | - University Of Strathclyde
|
|---|
| Sponsors | EPSRC (Engineering and Physical Sciences Research Council) |
|---|
| Supervisor | Bruce Stephen (Supervisor) & Blair David Brown (Supervisor) |
|---|