Enhanced backpropagation training algorithm for transient event identification

Conference · Fri Jan 01 04:00:00 EST 1993 · Transactions of the American Nuclear Society; (United States)

OSTI ID:7128932

Vitela, J; Reifman, J ^[1]

---

1. Argonne National Laboratory, IL (United States)

We present an enhanced backpropagation (BP) algorithm for training feedforward neural networks that avoids the undesirable premature saturation of the network output nodes and accelerates the training process even in cases where premature saturation is not present. When the standard BP algorithm is applied to train patterns of nuclear power plant (NPP) transients, the network output nodes often become prematurely saturated causing the already slow rate of convergence of the algorithm to become even slower. When premature saturation occurs, the gradient of the prediction error becomes very small, although the prediction error itself is still large, yielding negligible weight updates and hence no significant decrease in the prediction error until the eventual recovery of the output nodes from saturation. By defining the onset of premature saturation and systematically modifying the gradient of the prediction error at saturation, we developed an enhanced BP algorithm that is compared with the standard BP algorithm in training a network to identify NPP transients.

• OSTI ID: 7128932
• Report Number(s): CONF-931160--
• Conference Information: Journal Name: Transactions of the American Nuclear Society; (United States) Journal Volume: 69
• Country of Publication: United States
• Language: English

Related Subjects

22 GENERAL STUDIES OF NUCLEAR REACTORS
220400 -- Nuclear Reactor Technology-- Control Systems
220900* -- Nuclear Reactor Technology-- Reactor Safety
ACCIDENTS
ALGORITHMS
ARTIFICIAL INTELLIGENCE
DATA PROCESSING
DIAGNOSTIC TECHNIQUES
MATHEMATICAL LOGIC
MATHEMATICAL MODELS
NEURAL NETWORKS
PROCESSING
REACTOR ACCIDENTS
SATURATION
TRANSIENTS