Analysis and monitoring of a CANDU nuclear power plant using multivariate statistical process control methods

Abstract

<p>In the nuclear power industry, the ability to efficiently analyse historical data, and to detect and diagnose process faults in a timely manner are critical tasks in operating and maintaining a nuclear power plant. The objectives of this research were to prove that established Statistical Process Control (SPC) techniques could be used to analyse nuclear power plant data and to develop a hierarchical process monitoring methodology which could deliver relevant information to different functional groups within a plant. The use of established Multivariate SPC techniques to analyse nuclear power plant data was successfully proven in several areas and is considered a significant contribution to the development of data analysis tools for the nuclear energy industry. By analysing actual data from an operational nuclear power plant, the techniques were able to provide key insights into the process. Process tests and different plant configurations were easily identified. The multivariate techniques could relate the different plant configurations to sensor calibrations and process changes. Also, the techniques were able to identify two anomalies in the data which were not previously detected using the existing analysis tools. In order to produce a hierarchical process monitoring methodology, a multi-block, multi-level Principal Component Analysis algorithm, and associated prediction code, was developed and tested. This algorithm is an extension of existing multi-block, two-level algorithms and represents a contribution to the current state of Multivariate SPC techniques. It was found that the algorithm was very useful for analysis but marginal for delivering relevant information in a process monitoring capacity. This finding resulted in the third major contribution of this research work. The Multivariate SPC techniques are very useful for analysing nuclear power plant data but not as feasible for monitoring the process in an on-line manner. This was attributed to the goals defined for the monitoring methodology, the scaling method used for the data, and the numerous normal plant operating states.</p>