Two novel procedures for aggregating randomized model ensemble outcomes for robust signal reconstruction in nuclear power plants monitoring systems

Abstract

International audienceDetecting anomalies in sensors and reconstructing the correct values of the measured signals is of paramount importance for the safe and reliable operation of nuclear power plants. Auto-associative regression models can be used for the signal reconstruction task but in real applications the number of sensors signals may be too large to be handled effectively by one single model. In these cases, one may resort to an ensemble of reconstruction models, each one handling a small group of sensor signals; the outcomes of the individual models are then combined to produce the final reconstruction. In this work, three methods for aggregating the outcomes of a feature-randomized ensemble of Principal Components Analysis (PCA)-based regression models are analyzed and applied to two case studies concerning the reconstruction of 215 signals monitored at a Finnish nuclear Pressurized Water Reactor (PWR) and 920 simulated signals of the Swedish Forsmark-3 Boiling Water Reactor (BWR). Based on the insights gained, two novel aggregation procedures are developed for optimal signal reconstruction

Fault Detection in Nuclear Power Plants Components by a Combination of Statistical Methods

“…with overlapping, i.e., the same signal can belong to more than one group [12], [22]- [25], and without overlapping [26]- [28]. In practical applications, the latter strategy tends to be preferred because it allows for a smaller number of models to be developed, at a lower computational effort [28].…”

“…In practical applications, the latter strategy tends to be preferred because it allows for a smaller number of models to be developed, at a lower computational effort [28]. Two different approaches to grouping can be implemented: filter, and wrapper.…”

A belief function theory based approach to combining different representation of uncertainty in prognostics

“…With respect to i), we consider the average of the point predictors Other, more advanced techniques for ensemble aggregation of point predictions have been proposed in the literature, ranging from statistics methods like the mean and the median [7,30], to weighed averages of the model outcomes based on the global or local performances of the individual models [1,11]). Since our main objective in this work is the estimation of the prediction uncertainty, these techniques will be object of future research work.…”

“…Thus, techniques for the aggregation of point predictions, ranging from statistical methods, such as the mean and the median [7,30], to weighed average based on global or local performance measures of the individual models [1,11], are not considered in this work since they do not take into account uncertainty. Similarly, 4 techniques for the combination of individual probability distributions into a single aggregated probability distribution [12] cannot be used given that the two prognostic approaches provide different representations of the RUL uncertainty, i.e.…”

“…In practice, R=7 trajectories are used to train the prognostic models and the remaining trajectory to verify its performance. This leave-one-out procedure [7,29] has been repeated 8 times, using each time a different test trajectory.…”