Self-organizing neural networks for sequence representation and classification : application to the SCSS nuclear database

Author

Awatef Gacem

Date of Award

8-1992

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Electrical Engineering

Major Professor

Robert E. Uhrig, Joseph M. Googe

Committee Members

Dragana Brazakovic, Donald W. Bouldin

Abstract

A methodology based on self-organizing neural networks is introduced to compare events in nuclear power plants stored in the Sequence Coding Search System. This methodology uses a Self-organizing Feature Map (SOFM) to perform an unsupervised classification of the events. Because events are sequences of occurrences of variable length and occurrences are sequences of an arbitrary number of letters, the classification of events with a SOFM required the conversion of an event into a spatial representation.

To transform an event, a sequence of symbols of arbitrary length, into a fixed-size array of continuous values, an event was viewed as a sequence of overlapping triples, where a triple is a sequence of three occurrences. A linear combination of the numeric representations of the triples provided a spatial representation of an event.

A selectivity-based learning allowed a neural network to generate rapidly and without supervision a representation of the triples. By associating the symbols in an occurrence and integrating the occurrences of a triple into a pattern of activity of an array of units, this network converted automatically each triple into a distributed representation.

Because the spatial representation of events was distributed and reduced, the classification of the events was graded and focused on the major characteristics of an event without being overwhelmed by the micro features of its symbolic description. This novel representation of sequences of symbols allowed also the SOFM to process a large number sequences of symbols of arbitrary length. For the first time, the SOFM performed a classification of sequences which exploits the context of the entire sequence rather than a local-in-time context.

The clustering of events performed with the SOFM resulted in a Sequence Feature Map (SFM). The topological relationships among the clusters of events exhibited by the SFM provided clues to identify the unknown causes of event, and allowed the detection of unusual occurrences or events as well as the discovery of unsuspected relationships between events.

The new spatial representation for symbolic sequences of variable length has applications in many areas in which temporal sequences need to be processed using neural networks.

The use of self-organizing neural networks to represent and compare sequences of symbols allowed the discovery of a knowledge that is implicitly encoded in a large temporal database and therefore contributed to adding intelligent capabilities to databases when dealing with the tasks of data acquisition and database search.

Recommended Citation

Gacem, Awatef, "Self-organizing neural networks for sequence representation and classification : application to the SCSS nuclear database. " PhD diss., University of Tennessee, 1992.
https://trace.tennessee.edu/utk_graddiss/10885