Artificial Neural Networks Modeling as a Diagnostic and Decision Making Tool

Abstract

Fault diagnosis of a complex system, such as nuclear power plant (NPP), is an important part of the operator task. Artificial neural networks (ANN) model learns from the plant past behavior and is easy to develop if enough data is available for training, be it from actual plant history or from full-scale simulator fault situations. The operator is reluctant to depend on a “black-box” advice, and any reasonable explanation is welcome. Two examples of using ANN modeling for fault diagnosis are presented. The first is based on NPP full-scale simulator time history data of three instrument or component faults. The second example is the diagnosis of a seemingly erratic excessive steam turbine bearing vibration. The ANN models were able to correctly classes the causes of the faults. In the first case a knowledge extraction technique from trained ANN models, Causal Index, was used. It was able to provide plausible explanations for the ANN faults classification. In the second case, Genetic Algorithm was used as an optimization technique, with the ANN model providing the target function. It was able to identify the cause of the excessive vibration in a power plant turbine bearing.