Large scale nuclear sensor monitoring and diagnostics by means of an ensemble of regression models based on Evolving Clustering Methods

摘要

On-line sensor monitoring systems aim at detecting anomalies in sensors and reconstructing their correct signals during operation. Auto-associative regression models are usually adopted to perform the signal reconstruction task. In full scale implementations however, the number of sensors to be monitored is very large and cannot be handled effectively by a single reconstruction model. This paper tackles this issue by resorting to an ensemble of reconstruction models in which each model handles a small group of signals. In this view, firstly a procedure for generating the signal groups must be set. Then, a corresponding number of signal reconstruction models must be built on the bases of the groups and, finally, the outcomes of the reconstruction models must be aggregated. In this paper, three different signal grouping approaches are devised for comparison: pure-random, random-filter and random-wrapper. Signals are then reconstructed by Evolving Clustering Method (ECM) models. The median of the outcomes distribution is here retained as the ensemble aggregate. The ensemble approach is applied to a real case study concerning the validation and reconstruction of 792 signals measured at the Swedish boiling water reactor located in Oskarshamn.