Improved Explicability for Pump Diagnostics in Nuclear Power Plants

Abstract

To ensure the continued safe operation of many of the UK's fleet of advanced gas-cooled reactors (AGRs) effective and reliable monitoring of several of the key plant items is essential. Out of these key items, a significant portion of these are rotating plant assets, one asset in particular that is crucial to the operation of the station are the boiler feed pumps (BFPs). The BFPs in an AGR station move water from a condenser into a boiler, the water is then heated which produces steam and this steam turns the electricity-generating turbines. Currently, the operator of the AGR stations employs a time-based maintenance strategy for BFP assets: after a defined amount of time each asset is removed, replaced with a rotated spare, and a complete overhaul is then performed on the removed asset. This procedure can result in the removal of an asset before any significant wear has occurred, therefore increasing maintenance and generation costs. Conversely, this could result in an unplanned outage due to a component failure which leads to both a decrease in power output of the station and hence a decrease in revenue for the operator. Because these pumps are essential for the generation of electricity there are several pressure, temperature, vibration and speed parameters constantly monitored during the operation of this asset. Currently, data analysts have to manually analyse all this data by following a set diagnosis process, the consequential time burden on the analyst is therefore extremely high. Data-driven approaches to solve this problem, and other similar problems, have the capability to produce accurate results similar to what the analysts can achieve in a fraction of the time. However, the majority of these techniques are black box techniques and lack explicability which is often a requirement for problems involving critical assets in the nuclear industry. The main outcomes of this work are to address the time burden placed on the analysts by automating elements of the existing diagnosis process, through the implementation of an intelligent rule-based expert system, that provides adequate explicability to the user to satisfy requirements. Additionally, a recurring problem in the design of expert systems for industry is the cost involved with the knowledge elicitation process. Here we propose a questionnaire style approach, similar to what the domain experts currently use, to extract this knowledge without the need for a structured interview. By using this information a signal-to-symbol transformation algorithm is designed to assign time periods symbols that relate to the various rules defined by the domain experts. The final system combines the data-driven signal-to-symbol transformation algorithm and the rule-based expert system to produce a hybrid system that can be used to classify defects based on a set of rules and also explain to the user the reasoning behind this solution.