Data requirements for an anomaly detector in an automated safeguards system using neural networks.

摘要

An automated safeguards system must be able to detect and identify anomalous events in a near-real-time manner. Our approach to anomaly detection is based on the demonstrated ability of neural networks to model complex, nonlinear, real-time processes. By modeling the normal behavior of processes, we can detect how a system should behave and, thereby, detect when an abnormal state or event occurs. In this paper, we explore the computational intensity of training neural networks, and we discuss the issues involved in gathering and preprocessing the safeguards data necessary to train a neural network for anomaly detection. We explore data requirements for training neural networks and evaluate how different features of the training data affect the training and operation of the networks. We use actual process data to train our previous 3-tank model and compare the results to those achieved using simulated safeguards data. Comparisons are made on the basis of required training times in addition to correctness of prediction.