PNNL

Abstract

The Automated Radioxenon Analyzer/Sampler (ARSA) is a radioxenon gas collection and analysis system operating autonomously under computer control. The ARSA systems are deployed as part of an international network of sensors, with individual stations feeding radioxenon concentration data to a central data center. Because the ARSA instrument is complex and is often deployed in remote areas, it requires constant self-monitoring to verify that it is operating according to specifications. System performance monitoring is accomplished by over 200 internal sensors, with some values reported to the data center. Several sensors are designated as safety sensors that can automatically shut down the ARSA when unsafe conditions arise. In this case, the data center is advised of the shutdown and the cause, so that repairs may be initiated. The other sensors, called state of health (SOH) sensors, also provide valuable information on the functioning of the ARSA and are particularly useful for detecting impending malfunctions before they occur to avoid unscheduled shutdowns. Any of the sensor readings can be displayed by an ARSA Data Viewer, but interpretation of the data is difficult without specialized technical knowledge not routinely available at the data center. Therefore it would be advantageous to have sensor data automatically evaluated for the precursors of malfunctions and the results transmitted to the data center. Artificial Neural Networks (ANN) are a class of data analysis methods that have shown wide application to monitoring systems with large numbers of information inputs, such as the ARSA. In this work supervised and unsupervised ANN methods were applied to ARSA SOH data recording during normal operation of the instrument, and the ability of ANN methods to predict system state is presented.