Multivariate Diagnostics and Anomaly Detection for Nuclear Safeguards*

We first review recent literature that applies multivariate Shewhart and multivariate cumulative sum (Cusum) tests to detect anomalous data. These tests are used to evaluate residuals obtained from a simulated three-tank problem in which five variables (volume, density, and concentrations of uranium, plutonium, and nitric acid) in each tank are modeled and measured. We then present results from several simulations involving transfers between the tanks and between the tanks and the environment Residuals from a no-fault problem in which the measurements and model predictions are both correct are used to develop Cusum test parameters which are then used to test for faults for several simulated anomalous situations, such as an unknown leak or diversion of material from one of the tanks. The leak can be detected by comparing measurements, which estimate the true state of the tank system, with the model predictions, which estimate the state of the tank system as it \"should\" be. The no-fault simulation compares false alarm behavior for the various tests, whereas the anomalous problems allow us to compare the power of the various tests to detect faults under possible diversion scenarios. For comparison with the multivariate tests, univariate tests are also applied to the residuals.