Year
2006
Abstract
Safeguards data reviews of MBAs (Material Balance Areas) can be seen as a continuous verification process, broken down by the perioding reporting to guarantee the timeliness of inspections. The archive of review reports captures MBA history and its typical operation regularities. Most of these are known to the inspector through the experience s/he gained over the sequence of review exercises. To assist this process effectively, review tools also need to be aware of the MBA’s regularities, either ‘by being told’ (by the inspector) or ‘by learning them’ from the base of knowledge contained in review archives. For instance, in image reviews, inspectors tell change detection filters where to spot differences in gray-intensity by drawing areas of interest over the image plane. The focus of our work is rather on what review tools can learn in a semi-autonomous way from archives of past and closed reviews, as well as from the on-line compilation of review reports. Along this direction, we presented in previous work software tools that learn improved filtering rules by using the visual appearance of images reviewed by inspectors. In this paper, we argue that also the temporal succession of data should be modeled and considered by review tools. For example, in an MBA, flasks of nuclear material undergo a processing governed by a rather regular sequence of stages. The duration of each stage and the normal succession of processing steps are features that can be learnt. Specifically, we describe a model of flask processing in a MBA based on a hybrid Markov model (hMM). Markov models are discrete-state machines where state-to-state transitions have an associated probability. In our hybrid model, the relation between the observed quantities (e.g. gray-level change in specific image regions) and the underlying process is also dependent on a probabilistic function. A salient benefit of probability frameworks is their inferential capability of learning from historical evidence. Former inspectors’ reports can be employed to build a statistical description of the process, and to train the MM. The trained model keeps track of the state of processing and checks that the consequentiality of events is respected. Irregular event patterns are attributed a high probability of being false positives, and thus discarded from the review. In our experiments on real review data provided by inspectors, we verified that all significant images were still correctly selected by the hMM, whereas the false positives were reduced of a factor between 2.5 and 10 with respect to the methods currently in use.