A data-driven adaptive model-identification based large-scale sensor management system: Application to self powered neutron detectors

In this paper we propose an adaptive approach to manage large number of correlated sensors. Our approach is able to extract information (models) from these sensors that is relevant for performing fault diagnosis of these sensors. Such a situation involving large number of correlated sensors is encountered in large core nuclear reactors, for example. Since fault diagnosis methods are computationally intensive, it is helpful to organize the sensors into groups such that strongly correlated sensors belong to the same group. However, the groups/clusters need to be reorganized depending on operating conditions. We propose an adaptive method that is scalable to a large number of sensors and can adapt to changing operating conditions. Also, within each cluster, it is often required to adaptively rebuild new models/relations for sensors inside that cluster. We use the k-means algorithm for obtaining clusters and Principal Component Analysis (PCA) for finding relations between the sensors within a cluster. We demonstrate that significant speedup is achieved by parallelizing the various aspects of the above computation. A key requirement in managing a large number of sensors is the data and processing management. We demonstrate and compare a serial and parallel implementation of this method using SQLite for database management, Python for numerical computations, the Pycluster module for clustering and the Python multiprocessing module for code parallelization. The method is demonstrated for the above nuclear reactor application: with 140 sensors and 14,000 measurements for each sensor. The method turns out to scale very easily to such a large number. The implementation codes of our approach have been made available online. The utilized packages all being open source (FOSS) helps in the use of these codes in various safety critical applications which typically require complete verification/ratification. The cost saved due to the FOSS aspect of our implementation is another ad- antage.