Edf's plants monitoring through empirical modelling: Performance assessment and optimization

Abstract

Contact: redouane.seraoui@edf.fr EDF Group (Electricité de France) is one of Europe's leading energy players. With an installed capacity reaching the 127 GW (mainly nuclear, fossil, hydraulic), EDF was originally acting in France but is nowadays a major company also in the neighboring countries such as Germany, Italy or Great Britain. Most of EDF power plant components are either permanently or periodically monitored by specific techniques such as vibration, acoustic analysis, thermal imaging or oil analysis as an Equipment Condition Monitoring. Today, Condition Monitoring based on statistical modelling algorithms using process data is becoming more prevalent in the power, chemical and aerospace industries. The main interest for EDF is to complete the specific techniques and to improve the detectability of the faults in order to have time to plan maintenance actions. All these statistical methods use the same principle for fault detection, which is to compare observed data to estimated data: A difference will reveal the presence of an anomaly, which can be related either to equipment or instrumentation. EDF R & D division has evaluated for its internal clients off-the-shelf monitoring tools, which embed such methods and has implemented auto associative kernel regression (AAKR) and evolving clustering method (ECM) on a Matlab platform. Up to now, the design of the industrial monitoring tools has centred about ease of use, allowing the user to gain access to complex analysis methods without great effort. However, for a proper engineering modelling, one must pay greater attention to the issue of parameters tuning. For that purpose, the R & D division is writing a guideline tackling the main issues entailed by learning-based modelling in a fleet-wide monitoring context. The paper gives first an overview of the theoretical principles behind AAKR and ECM then it presents usual criteria in order to quantify three important aspects of a model performance: the accuracy, the robustness and the ability to isolate faults called spill-over. The performance of the two methods is compared regarding simulated faults on real plant data. Secondly, the sensitivity of these methods has been investigated, particularly regarding the type of kernel, the maximum radius of a cluster, the number of samples in the training, the sensors weighting in the distance function or the number of signals in the model. Finally, an approach of model performance optimization based on these sensitivity indicators has been assessed. This procedure helps to identify the faulty sensors 1 adjust their weights in the overall distance and thus to avoid spill-over and increase the model robustness.